Biblioteca Digital | FCEN-UBA | Waissbein, Ariel. 2013 "Algoritmos

Transcription

Algoritmos de deformación para la resolución de
sistemas polinomiales
Waissbein, Ariel
2013
Tesis Doctoral
Facultad de Ciencias Exactas y Naturales
Universidad de Buenos Aires
www.digital.bl.fcen.uba.ar
Contacto: digital@bl.fcen.uba.ar
Este documento forma parte de la colección de tesis doctorales y de maestría de la Biblioteca
Central Dr. Luis Federico Leloir. Su utilización debe ser acompañada por la cita bibliográfica con
reconocimiento de la fuente.
This document is part of the doctoral theses collection of the Central Library Dr. Luis Federico Leloir.
It should be used accompanied by the corresponding citation acknowledging the source.
Fuente / source:
Biblioteca Digital de la Facultad de Ciencias Exactas y Naturales - Universidad de Buenos Aires
UNIVERSIDAD DE BUENOS AIRES
Facultad de Ciencias Exactas y Naturales
Departamento de Matemática
Algoritmos de deformación para la resolución de
sistemas polinomiales
Tesis presentada para optar al título de Doctor de la Universidad de
Buenos Aires en el área Ciencias Matemáticas
Ariel Waissbein
Director de tesis: Dr. Guillermo Matera
Consejero de estudios: Dr. Guillermo Matera
Buenos Aires, 17 de Diciembre de 2013
Algoritmos de deformación para la resolución
de sistemas polinomiales
Esta tesis está dedicada a ciertas tareas computacionales de geometría
algebraica en característica cero. Apuntamos a analizar y descubrir la complejidad de problemas definidos por sistemas de ecuaciones polinomiales con
una perspectiva de álgebra computacional. La intratabilidad computacional
de los enfoques generalistas a los problemas de geometría computacional nos
impele a estudiar familias particulares de sistemas de ecuaciones polinomiales
en los que la complejidad del peor caso es tratable (y significativamente más
baja que la del caso general). Cuando sea posible, proveeremos un método
eficiente para encontrar su solución.
Como “brújula” para determinar estas familias usamos técnicas de deformación las que, según mostraremos, son sensibles a problemas con buenas
propiedades semánticas. Entonces, este trabajo consiste en establecer algunos problemas de eliminación que son tratables y exhibir algoritmos eficientes
que los resuelven.
Nuestras técnicas de deformación se basan en un procedimiento de levantamiento à la Newton–Hensel que se adapta bien para producir algoritmos
que corren en menos pasos cuando las propiedades semánticas referenciadas
anteriormente son buenas. Construiremos, entonces, un catálogo de resultados sobre la resolución de sistemas de ecuaciones polinomiales, usando algoritmos de álgebra altamente eficientes, que constituyen mejoras en relación
con el estado del arte.
Palabras clave: algoritmos eficientes, ecuaciones polinomiales, eliminación
geométrica, levantamiento de Newton-Hensel, algoritmos simbólicos, algoritmos probabilísticos, complejidad.
Deformation Algorithms for Polynomial
System Solving
This thesis is devoted to computational tasks of basic algebraic geometry
in characteristic zero. We aim to analyse and discover the complexity of problems defined by systems of polynomial equations from a computer algebra
perspective. The computational intractability of a general approach to geometric elimination problems compels us to study the difficulty of elimination
for particular families of polynomial equation systems where worst-case complexity is tractable (and significantly lower than the complexity of tackling
the general case). When possible, we provide an efficient solution method.
As our “compass” for determining these families, we use deformation techniques which, we will show, are susceptible to problems with well-posed semantic properties. Hence, this work consists in establishing some elimination
problems that are tractable, and for these, exhibiting efficient algorithms that
tackle them.
Our deformation techniques rely on a Newton-Hensel lifting procedure
which adapts well in order to obtain algorithms running in fewer steps when
certain semantic parameters are “low”. Using highly-efficient algorithms for
constructing these geometric elimination procedures, we develop a catalogue
of results on polynomial system solving that improve over the prior art.
Keywords: efficient algorithms, polynomial equations, geometric elimination, Newton-Hensel lifting, symbolic algorithms, probabilistic algorithms,
complexity.
Agradecimientos
Quisiera agradecer de manera especial y sincera al profesor Dr. Guillermo
Matera por aceptar ser mi director y acompañarme durante este doctorado
con empeño y generosidad, contribuyendo con un aporte invaluable.
Quiero también expresar mi agradecimiento al profesor Dr. Joos Heintz
por haber generado ese interés en mi por la eliminación geométrica, por las
muchas charlas matemáticas y de tantas otras cosas.
Me gustaría agradecer a los varios profesores y alumnos, co-autores, conferencistas y aquellos otros científicos con los que me crucé durante esta
carrera.
Debo agradecer a mi familia por su apoyo, por ayudarme a destinar las
horas necesarias de trabajo a este doctorado, y por las palabras de aliento.
6
Contents
1. Introduction
23
1.1. A catalogue . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.1.1. Generalized Pham Systems . . . . . . . . . . . . . . . . 28
1.1.2. Lifting from ramified fibres . . . . . . . . . . . . . . . . 30
1.1.3. Sparse systems . . . . . . . . . . . . . . . . . . . . . . 31
2. Preliminaries
35
2.1. Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.2. Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.2.1. Geometric Solutions . . . . . . . . . . . . . . . . . . . 40
2.3. Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.3.1. Complexity model
. . . . . . . . . . . . . . . . . . . . 42
2.3.2. Probabilistic identity testing . . . . . . . . . . . . . . . 43
2.3.3. Basic complexity estimates . . . . . . . . . . . . . . . . 43
2.4. Deformation algorithms
. . . . . . . . . . . . . . . . . . . . . 44
2.4.1. The lifting process . . . . . . . . . . . . . . . . . . . . 46
2.4.2. From minimal equations to geometric solutions
3. Generalized Pham Systems
. . . . 53
57
3.1. A catalogue of generalized Pham systems . . . . . . . . . . . . 60
7
8
CONTENTS
3.1.1. Pham Systems . . . . . . . . . . . . . . . . . . . . . . 60
3.1.2. Systems Coming from a Semidiscretization of certain
Parabolic Differential Equations . . . . . . . . . . . . . 60
3.1.3. Reimer Systems . . . . . . . . . . . . . . . . . . . . . . 61
3.1.4. Generalized Pham Systems . . . . . . . . . . . . . . . . 61
3.2. The Solution of a Generalized Pham System . . . . . . . . . . 62
3.2.1. The architecture of our solution method . . . . . . . . 63
3.2.2. Designing suitable deformations . . . . . . . . . . . . . 64
3.2.3. Preparation: random choices . . . . . . . . . . . . . . . 68
3.2.4. The main algorithm
. . . . . . . . . . . . . . . . . . . 74
4. Deformations for ramified fibres
81
4.1. Puiseux expansions of space curves . . . . . . . . . . . . . . . 82
4.2. Lifting procedures for ramified fibres . . . . . . . . . . . . . . 84
4.2.1. Properties of the ideal I (ℓ,k) . . . . . . . . . . . . . . . 86
4.2.2. The unramifiedness of the morphism π (ℓ,k) at T = 0 . . 90
4.2.3. A global Newton–Hensel lifting . . . . . . . . . . . . . 94
4.2.4. Unramifiedness and flatness conditions . . . . . . . . . 96
4.3. Algorithms and complexity estimates . . . . . . . . . . . . . . 99
4.4. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.4.1. Pham–Brieskorn systems . . . . . . . . . . . . . . . . . 107
4.4.2. Systems coming from a semidiscretization of certain
parabolic differential equations . . . . . . . . . . . . . 107
4.4.3. A common approach to both examples . . . . . . . . . 109
4.4.4. Reimer Systems . . . . . . . . . . . . . . . . . . . . . . 113
5. Deformations for Sparse Systems
119
5.1. Sparse Elimination . . . . . . . . . . . . . . . . . . . . . . . . 123
9
CONTENTS
5.1.1. Mixed subdivisions . . . . . . . . . . . . . . . . . . . . 125
5.1.2. Degree estimates in the sparse setting . . . . . . . . . . 128
5.2. Solution of a generic sparse system . . . . . . . . . . . . . . . 131
5.2.1. The polyhedral deformation . . . . . . . . . . . . . . . 132
5.2.2. On the genericity of the initial system
. . . . . . . . . 135
5.2.3. Outline of the algorithm . . . . . . . . . . . . . . . . . 140
5.2.4. Geometric solutions of the starting varieties . . . . . . 142
5.2.5. A geometric solution of the curve Vb . . . . . . . . . . . 148
5.2.6. Solving the generic sparse system . . . . . . . . . . . . 165
5.3. The solution of the input system
. . . . . . . . . . . . . . . . 166
5.3.1. Solution of the second deformation . . . . . . . . . . . 167
5.3.2. Solving the input system . . . . . . . . . . . . . . . . . 168
Bibliography
171
10
CONTENTS
Introducción
Dado un sistema de ecuaciones polinomiales sobre los racionales queremos
contestar preguntas sobre sus soluciones. Según la tradición en informática,
los problemas se especifican por sus parámetros sintácticos. En el caso que
nos compete, esto significa buscar un algoritmo (no necesariamente uniforme) que, dados enteros d, n y s, calcula las soluciones de un sistema de s
polinomios n-variados de grado total acotado por d cuando éste define una
variedad de dimensión cero. Es sabido que este problema es P # -duro y que
el problema de decidir si un sistema definido sobre los enteros define o no la
variedad vacía es NP- y NPC -duro ([HM93], [SS93a]).
De hecho, tanto el acercamiento simbólico como el numérico a la resolución de sistemas de ecuaciones polinomiales se hace por demás intrincado.
Los métodos numéricos no pueden ser aplicados directamente a sistemas paramétricos, sobre-determinados, sub-determinados o degenerados —que son
de gran interés (ver, e.g., [Par95]). Más aún, cuando comparamos los acercamientos simbólicos con los numéricos de [SS93a], [SS93b], [SS93c], [SS96b],
[SS94], [CS99] y [BCSS98], es posible mostrar que este acercamiento numérico
es inferior en términos de complejidad bit (ver [CHMP01] y [CJPS02]).
Por otro lado, el famoso resultado de Mayr y Meyer ([MM82]) implica
que el “Hauptproblem der Idealtheorie” (i.e., el problema de pertenencia a
un ideal) es EXPSPACE-completo; y, por ende, esto es evidencia de que
los métodos simbólicos basados en re-escritura, tales como aquellos basados
en bases de Gröbner (ver, e.g., [Buc85]), necesitan de un espacio de memoria exponencial en el peor de los casos (ver también, e.g., [May89], [KM96]
y [Küh98]). De otro resultado, éste de D. Lazard, T. Mora, W. Masser y
P. Philippon (ver [Bro87]), podemos deducir que el número de operaciones
aritméticas necesarias para resolver un sistema polinomial expresado por su
codificación densa es inevitablemente exponencial. Sin embargo, los métodos
11
12
CONTENTS
simbólicos basados en re-escritura son los preferidos en la gran mayoría del
software disponible.
Nosotros proponemos evitar este crecimiento de la complejidad para algoritmos simbólicos restringiéndonos a “preguntas geométricas” o a casos
especiales de sistemas de ecuaciones polinomiales, e.g., aquellos con un número finito de soluciones. Sin embargo, incluso para sistemas con un conjunto
finito de soluciones, la cantidad de operaciones aritméticas que necesitan al2
goritmos simbólicos optimales es del orden de sdO(n ) . Este hecho se debe a
que los algoritmos simbólicos están limitados por las representaciones densas (o ralas) de los polinomios de entrada. Podemos evitar este dilema si la
representación de los polinomios de entrada está dada por cajas negras o
“(division-free) straight-line programs” que los evalúan. Aún en este caso, la
cota para el peor caso se limita a sdO(n) para aquellos algoritmos que fueron
implementados usando las reglas de ingeniería de software (ver [HKR11] y
[HKR12]).
Un catálogo
Tomamos el problema de buscar clases de sistemas de ecuaciones que
pueden resolverse con recursos computacionales “razonables” y disponibles
para ingenieros y científicos en nuestros días.
Nuestro acercamiento se construye sobre los cimientos de [GHM+ 98],
[GHH+ 97], [GHMP97] y [Par95]. En estos artículos los autores introducen
algoritmos para resolución de sistemas polinomiales basándose en una deformación homotópica que es “seguida” mediante un levantamiento à la NewtonHensel. Dichos algoritmos toman como entrada un straight-line program que
calcula los polinomios que definen el sistema bajo estudio y devuelven una
solución geométrica de este sistema en un tiempo polinomial en un parámetro
que llaman el grado del sistema, y que está acotado por el número de Bézout
del sistema.
Además, también nos basamos en las enseñanzas de [HKP+ 00], [GLS01],
[HMW01] y [Sch03]. En esta segunda familia de artículos, los autores aíslan
un algoritmo de deformación basado en la técnica de levantamiento anterior, estiman el costo de este algoritmo en términos de parámetros más finos
de carácter geométrico y producen procedimientos eficientes que calculan la
CONTENTS
13
solución de algunos sistemas polinomiales.
Antes de describir estas técnicas de deformación queremos dar un paso
atrás y recordar un resultado de [CGH+ 03] (cf. [HMPW98], [Par00], [GH01],
[BP06]), que muestra que cualquier algoritmo universal y robusto que resuelve
ciertos sistemas polinomiales sobre los complejos requiere de al menos DΩ(1)
operaciones aritméticas, donde D denota el número de Bézout del sistema de
entrada. Ergo, para los sistemas polinomiales en nuestro catálogo apuntamos
a: a) reemplazar a D por parámetros más finos que reflejen las propiedades
geométricas del sistema, y b) reemplazar a Ω(1) en la reciente estimación por
un número que sea lo más bajo posible. Nuestro esfuerzo no constituye un
ataque al caso general, sino —como ya lo explicamos— para que esto quede
reflejado en los casos particulares que forman parte de nuestro catálogo.
Supongamos dados polinomios f1 , . . . , fn ∈ Q[X1 , . . . , Xn ] que definen un
sistema de dimensión cero V ⊂ Cn . Supongamos que podemos definir una
curva algebraica V ⊂ Cn+1 y un morfismo dominante y genéricamente no ramificado π : V → C tal que vale π −1 (1) = {1} × V , donde V es el conjunto de
ceros comunes en Cn+1 de ciertos polinomios F1 , . . . , Fn ∈ Q[T, X1 , . . . , Xn ]
y el morfismo π está dado por la regla π(t, x1 , . . . , xn ) := t. Entonces, tomando como entrada una descripción completa de una fibra no ramificada
π −1 (t0 ) (ver la Sección 2.2.1 para una definición precisa), podemos calcular
una descripción de cualquier fibra π −1 (t), y por lo tanto una de V . Llamamos
“algoritmo de levantamiento” a un procedimiento que resuelve este problema.
Típicamente produciremos a F1 , . . . , Fn como perturbaciones de f1 , . . . , fn
de manera que el sistema cero-dimensional {F1 (t0 , X) = 0, . . . , Fn (t0 , X) =
0} es “fácil” de resolver y vale Fi (1, X) = fi (X) para 1 ≤ i ≤ n. Para
este algoritmo de levantamiento, pedimos que los polinomios F1 , . . . , Fn de
Q[T, X] := Q[T, X1 , . . . , Xn ] formen una sucesión regular y generen un ideal
radical en Q[T, X]. También pedimos un punto t0 ∈ Q tal que su fibra por π
sea no ramificada, y una forma lineal U ∈ Q[X] tal que U separe los puntos
de π −1 (t0 ).
El algoritmo de levantamiento calcula una descripción completa de V a
partir de un straight-line program en Q[T, X] que evalúa a F1 , . . . , Fn y una
descripción completa de la fibra π −1 (t0 ). La salida del algoritmo consiste en
n+1 polinomios univariados que forman la codificación densa de una solución
geométrica.
14
CONTENTS
Con nuestras hipótesis, existen #π −1 (t0 ) n-uplas de series formales R :=
(R1 , . . . , Rn ) ∈ C[[T − t0 ]]n que son solución del sistema F1 = 0, . . . , Fn = 0,
i.e., vale que Fi (T, R) = 0 en C[[T − t0 ]] para 1 ≤ i ≤ n. El levantamiento de
Newton-Hensel que mencionamos se usa para aproximar estas series formales
de las cuales se calcula la descripción de V.
Seguidamente ilustramos la técnica de levantamiento mostrando cómo es
que el operador formal de Newton-Hensel aproxima a las series formales antes
referidas. Esto no descubre, de ninguna manera, los pasos que seguimos en
nuestros algoritmos y que detallaremos en las próximas secciones.
Sea (t0 , ξ) ∈ Cn+1 un punto en la fibra π −1 (t0 ). Entonces, de nuestras
hipótesis se sigue que existe una única n–upla de series formales R(ξ) :=
(ξ)
(ξ)
(R1 , . . . , Rn ) ∈ C[[T − t0 ]]n tal que R(ξ) (0) = ξ y Fi (T, R(ξ) (T )) = 0 en
C[[T − t0 ]] para cada 1 ≤ i ≤ n .
Denotemos por JF (X) := (∂Fi /∂Xj )1≤i,j≤n a la matriz Jacobiana de
F1 , . . . , Fn con respecto a X1 , . . . , Xn y sea


 
F1 (T, X)
X1


 
..
NF (X) :=  ...  − (JF (T, X))−1 

.
Fn (T, X)
Xn
el operador (formal) de Newton-Hensel asociado al sistema. Para cada κ ∈
Z≥0 , sea
(ξ,κ)
R(ξ,κ) := (R1 , . . . , Rn(ξ,κ) ) := NFκ (T, ξ)
la κ–ésima iteración del operador de Newton-Hensel NF comenzando desde
ξ. Entonces, afirmamos que:
det(JF (T, R(ξ,κ) )) ∈
/ (T − t0 )C[T ](T −t0 ) , y
κ
Fi (T, R(ξ,κ) ) ∈ (T − t0 )2 C[T ](T −t0 ) para 1 ≤ i ≤ n.
Más aún, también
Los primeros 2κ términos en (T − t0 ) de R(ξ,κ+1) y R(ξ,κ) son iguales,
κ
(ξ,κ+1)
(ξ,κ)
i.e., Ri
− Ri
∈ (T − t0 )2 C[T ](T −t0 ) para 1 ≤ i ≤ n.
Es fácil ver que estas afirmaciones son verdaderas para κ = 0, porque por
hipótesis el ideal (F1 (t0 , X), . . . , Fn (t0 , X)) de C[X] es radical, ya que la fibra
15
CONTENTS
π −1 (t0 ) se supone no ramificada. Esto implica que det(JF (t0 , ξ)) 6= 0 y luego
det(JF (T, R(ξ,0) )) ∈
/ (T − t0 )C[T ](T −t0 ) . Por otro lado, vale que Fi (T, R(ξ,0) ) =
Fi (T, ξ) ∈ (T − t0 )C[T ](T −t0 ) para 1 ≤ i ≤ n. De la identidad
R(ξ,1) = R(ξ,0) + JF (T, R(ξ,0) )−1 F (T, R(ξ,0) )
(ξ,1)
deducimos que Ri
(ξ,0)
− Ri
∈ (T − t0 )C[T ](T −t0 ) para 1 ≤ i ≤ n.
El paso inductivo se sigue de la definición del operador NF y el desarrollo
de Taylor de Fi y det(JF ) como series formales en (T − t0 ): si para cada
1 ≤ i ≤ n, multiplicamos cada miembro de la igualdad


F1 (T, R(ξ,κ) )


..
R(ξ,κ+1) − R(ξ,κ) = −JF−1 (T, R(ξ,κ) ) · 

.
(ξ,κ)
Fn (T, R
)
por la i–ésima fila de JF (T, R(ξ,κ) ), se sigue que
JF (T, R(ξ,κ) )i · (R(ξ,κ+1) − R(ξ,κ) ) = −Fi (T, R(ξ,κ) ).
(1)
Combinando (1) con la congruencia
Fi (T, R
(ξ,κ+1)
) ≡ Fi (T, R
(ξ,κ)
n
X
∂Fi
(ξ,κ+1)
(ξ,κ)
(T, R(ξ,κ) ) · (Rj
− Rj )
)+
∂Xj
j=1
mod (R(ξ,κ+1) − R(ξ,κ) )2 de C[[T − t0 ]](T −t0 ) , se deduce que Fi (T, R(ξ,κ+1) ) ≡ 0
mod (R(ξ,κ+1) − R(ξ,κ) )2 in C[[T − t0 ]](T −t0 ) . De la hipótesis inductiva se sigue
κ
κ+1
que (R(ξ,κ+1) −R(ξ,κ) )2 ⊂ ((T −t0 )2 )2 ⊂ (T −t0 )2 lo que prueba la segunda
parte del paso inductivo.
La última parte del paso inductivo sale de la siguiente igualdad


F1 (T, R(ξ,κ+1) )


..
R(ξ,κ+2) − R(ξ,κ+1) = −JF−1 (T, R(ξ,κ+1) ) · 

.
(ξ,κ+1)
Fn (T, R
)
y lo que acabamos de probar. Por otro lado, del desarrollo del determinante
Jacobiano det(JF ) se puede inferir que
det(JF (T, R(ξ,κ+1) ) ≡ det(JF (T, R(ξ,κ) ) +
n
X
∂ det(JF )
(ξ,κ+1)
(ξ,κ)
(T, R(ξ,κ) ) · (Rj
− Rj )
∂X
j
j=1
16
CONTENTS
mod (R(ξ,κ+1) − R(ξ,κ) )2 en C[[T − t0 ]](T −t0 ) . Así, de la primera parte de la
hipótesis inductiva deducimos su contraparte en el paso inductivo.
En el transcurso de esta tesis produciremos distintos algoritmos de levantamiento cuya algorítmica no puede inferirse directamente de los razonamientos anteriores, que se basan en contribuciones a la algorítmica de [GLS01] y
[Sch03], y usan aproximadamente O(deg V deg π) operaciones aritméticas en
Q (omitiendo algunos términos de orden poli-logarítmico en los parámetros
de la expresión), donde deg V es el grado de la variedad V y deg π el grado
del morfismo π (ver la Sección 2.3 para las definiciones de medidas de complejidad y la Sección 2.4 para obtener más precisiones sobre este resultado).
Vale aclarar que en nuestro caso deg π ≤ deg V ≤ D, donde D es el número
de Bézout del sistema que define a V.
Teniendo a este algoritmo disponible, el punto crítico para la aplicación
del algoritmo de homotopía/levantamiento es el de obtener morfismos π con
“grado bajo” y una fibra no ramificada que sea “fácil de resolver”.
Sistemas generalizados de Pham
Comenzamos nuestro catálogo con una clase importante de sistemas cuadrados, con coeficientes racionales, que definen una variedad cero-dimensional
sobre los complejos, llamados sistemas generalizados de Pham (ver [PS04] y
[DMW09]) o intersecciones completas estrictas (ver [CDS96]), que aparecen
relacionados con distintos problemas en geometría algebraica computacional (ver, e.g., [MT00], [HM00]). Un sistema generalizado de Pham puede ser
descripto someramente como el resultado de la deformación de una singularidad proyectiva de intersección completa; intuitivamente se corresponde a
la noción de un sistema “sin puntos en el infinito” (ver [PS04, Remark 17] o
[CDS96, Section 1]).
Un sistema n–dimensional generalizado de Pham está definido por n polinomios de la forma
φ1 + ϕ1 , . . . , φ n + ϕn ,
donde para cada 1 ≤ i ≤ n vale que φi ∈ Q[X1 , . . . , Xn ] es homogéneo, ϕi
es un polinomio (no necesariamente homogéneo) de grado total menor que
di := deg φi , y tal que φ1 , . . . , φn definen la variedad proyectiva vacía de
Pn−1 (C).
17
CONTENTS
Un ejemplo específico de sistema generalizado de Pham es el de un sistema
(n–dimensional) de Pham, que queda definido por polinomios de la forma
X1d1 + ϕ1 , . . . , Xndn + ϕn ,
donde ϕi ∈ Q[X1 , . . . , Xn ] es de grado menor que di para 1 ≤ i ≤ n (en la Sección 3.1 esgrimiremos otros ejemplos interesantes de sistemas generalizados
de Pham).
Las soluciones de un sistema de Pham f1 = 0, . . . , fn = 0 dado por
polinomios en Q[X1 , . . . , Xn ] pueden calcularse con una adaptación de [Sch03]
o incluso usando un resultado de [BMWW04, Section 5] —trabajo del que
soy co-autor. En breve, podríamos aplicar el algoritmo de proyección de los
artículos citados a la deformación definida por el morfismo πW : W → A1
dado por la regla πW (t, x) := t, donde la variedad W ⊂ An+1 viene dada por
las soluciones del sistema
ai Xidi + T (fi − ai Xidi ) + bi (1 − T ) = 0 (1 ≤ i ≤ n),
siendo ai el coeficiente de X di en fi y bi un racional arbitrario elegido aleatoriamente 1 ≤ i ≤ n. En [BMWW04] exhibimos un sistema que resuelve
sistemas de Pham con complejidad cuadrática y usa un algoritmo de homotopía directamente (para el caso d1 = . . . = dn ). Describiremos dicho
procedimiento más adelante en el Capítulo 4 en un contexto distinto.
Desafortunadamente, el anillo de coordenadas de un sistema generalizado
de Pham carece de la estructura monomial simple que aparece en los sistemas
de Pham, y eso hace que este procedimiento (cf. [BMWW04]) —pero también los de [MP97], [GLGV98], [MP00], [MT00]— devengan en algoritmos con
complejidad más que cuadrática en el número de Bézout del sistema de entrada. A saber, mirando el producto deg(πW ) deg W, que es el término dominante en la estimación de complejidad de nuestro algoritmo, para este caso particular vale que deg(πW ) = D := d1 · · · dn y deg W ≤ E := (d1 + 1) · · · (dn + 1)
por la desigualdad de Bézout. Entonces, cuando n ≫ máx{d1 , . . . , dn } se
sigue que E ≫ D y luego DE ≫ D2 .
No obstante, veremos que podemos producir un algoritmo probabilístico
que resuelve sistemas generalizados de Pham en complejidad cuadrática en el
número de Bézout de la entrada. Dicho algoritmo hará uso de, no una, sino
n + 1 homotopías sobre curvas producidas artificialmente, a fin de computar
la solución del sistema original.
18
CONTENTS
En el Capítulo 3 nos abocaremos al diseño de dicho algoritmo y la prueba
de los resultados matemáticos necesarios. Este es uno de los dos resultados
principales de [DMW09]; artículo del que soy co-autor. El segundo resultado
de aquel artículo dice que la resolución de sistemas generalizados de Pham
por algoritmos robustos y universales necesita de al menos DΩ(1) operaciones
aritméticas en Q (en el peor caso), donde D es el número de Bézout del sistema generalizado de Pham. Esta cota es independiente de la representación
de entrada y salida del algoritmo, aunque el exponente en la Ω sí depende
de dicha representación. Por ejemplo, mostramos que si usáramos la representación densa o rala (por rala entendemos la lista de todos los coeficientes
no nulos), entonces el exponente de D en la cota de complejidad es del orden
de Ω(D), mientras que si usamos un straight–line program para representarlos entonces el exponente es del orden de Ω(D1/2 ). Este segundo resultado
implica que nuestra cota superior es casi optimal.
Levantamiento de fibras ramificadas
El segundo ítem del catálogo es una generalización del algoritmo de levantamiento que presentamos y usamos arriba. Sea V ⊂ Cn+1 una curva
algebraica definida sobre Q, y supongamos que el morfismo π : V → C inducido por la proyección canónica en la primera coordenada es dominante
y genéricamente no ramificado. Supongamos dada la estructura infinitesimal
de una fibra finita y ramificada π −1 (tr ). Queremos calcular una descripción
completa de una fibra arbitraria π −1 (t). Mientras que en el caso de una fibra no ramificada podíamos usar un levantamiento à la Newton-Hensel para
aproximar las soluciones y derivar una descripción completa de una fibra
arbitraria π −1 (t), esto se vuelve imposible en el caso de que la fibra sea ramificada. A saber, supongamos dada una fibra π −1 (tu ) no ramificada para
cierto tu ∈ Q. Recordemos que el cuerpo de fracciones C((T − tu )) del anillo
de series formales C[[T − tu ]] no es algebraicamente cerrado, y que el cuerpo
de series de Puiseux Q(T − tu )∗ es una clausura algebraica de Q((T − tu ))
(ver, e.g., [Wal50]). Por lo expuesto anteriormente (Sección “Un catálogo”) se
sigue que todas las soluciones en An (C(T − tu )∗ ) del sistema
F1 (T, X) = 0, . . . , Fn (T, X) = 0
(2)
están, de hecho, en C[[T − tu ]]n . Sin embargo, cuando consideramos a (2)
como un sistema en An (C(T − tr )∗ ), no podemos aplicar el argumento de la
19
CONTENTS
Sección “Un catálogo”; de hecho, sucede que no todas las soluciones están en
C[[T − tr ]]n . En particular, no podemos usar el algoritmo de levantamiento
para este caso.
Una manera de salvar este problema es requiriendo una descripción completa del conjunto de partes singulares de los desarrollos en series de Puiseux
de las ramas de V que están sobre tr (ver la Sección 4.1 para más detalles). En
el Capítulo 4 exhibiremos un algoritmo que calcula una descripción completa
de una fibra arbitraria π −1 (t), dada dicha descripción de una fibra ramificada,
y que usa aproximadamente O(deg V(deg π)α ) operaciones aritméticas en Q,
siendo α = 1 en varios ejemplos importantes. Este es el resultado principal
de [BMWW04], del que soy co-autor, que extiende y mejora los resultados
de [HKP+ 00] y [Sch03] (cf. [DMW09]).
Asimismo repasaremos algunos ejemplos de familias a las que le subyace
una fibra ramificada que es “fácil” de resolver. Para esos casos, nuestro nuevo algoritmo de levantamiento calculará una solución completa de cualquier
miembro de la familia V.
Por ejemplo, sean n, d ∈ N y consideremos el sistema de Pham
f1 := X1d − ϕ1 (X), . . . , fn := Xnd − ϕn (X),
la curva definida por el sistema
F1 := X1d + T ϕ1 (X), . . . , Fn := Xnd + T ϕn (X),
y el morfismo π dado por π(t, x) := t. Entonces, la fibra π −1 (0) = {0} tiene
un solo punto y es ramificada. En este caso, las partes singulares de las ramas
de la curva {F1 = 0, . . . , Fn = 0} que están sobre T = 0 son
1/d
(ξ i1 α1 T 1/d , . . . , ξ in αn1/d T 1/d )
para 0 ≤ i1 , . . . , in ≤ d − 1, donde α := (α1 , . . . , αn ) := (ϕ1 (0), . . . , ϕn (0)).
Veremos que es posible aplicar el nuevo algoritmo de levantamiento para
resolver este sistema a partir de la fibra de t = 0 en aproximadamente
O(T deg V deg π) = O(Td2n ) operaciones aritméticas en Q.
La familia de sistemas de ecuaciones del próximo ítem es también un
ejemplo donde podemos aplicar este nuevo algoritmo.
20
CONTENTS
Sistemas ralos y homotopías poliedrales
Cerramos nuestro catálogo con un procedimiento para calcular todas las
soluciones de un sistema polinomial ralo (en el que el conjunto de coeficientes
no nulos es pequeño) y que se vale de una homotopía poliedral.
El famoso resultado de D.N. Bernstein, A.G. Kushnirenko y A.G. Khovanski ([Ber75], [Kus76], [Kho78]) acota el número de soluciones de un sistema polinomial en términos de un invariante combinatorio asociado a los
exponentes de los monomios con coeficientes no nulos de los polinomios del
sistema. Explícitamente, el Teorema de Bernstein-Kushnirenko-Khovanski
(que abreviamos por BKK) dice que el número de soluciones aisladas en
el toro complejo (C∗ )n de un sistema polinomial dado por n ecuaciones en n
variables está acotado por el volumen mixto de los polítopos de Newton de
los polinomios que definen al sistema.
Los métodos numéricos (de continuación homotópica) para sistemas ralos se basan típicamente en una familia de deformaciones que se llaman homotopías poliedrales ([HS95], [VVC94], [VGC96], [Roj03]). Las homotopías
poliedrales preservan el polítopo de Newton del sistema polinomial de entrada; asimismo, les subyace una versión efectiva del Teorema BKK (ver, e.g.,
[HS95], [HS97]).
Supongamos dado un sistema cero-dimensional (∆1 , . . . , ∆n )-ralo dado
por n polinomios f1 , . . . , fn ∈ Q[X1 , . . . , Xn ], con soportes ∆1 , . . . , ∆n ⊂ Zn
respectivamente. Sea V ⊂ (C∗ )n la variedad definida por los ceros comunes de
f1 , . . . , fn en (C∗ )n . Entonces, una homotopía poliedral consiste en una curva
algebraica V ⊂ (C∗ )n+1 tal que la proyección π : V → C∗ , π(t, x) := t en la
primera coordenada es dominante, con fibras genéricamente no ramificadas
cuyo grado es igual al volumen mixto M V (conv(∆1 ), . . . , conv(∆n )) de las
cápsulas convexas de ∆1 , . . . , ∆n , tal que vale la identidad π −1 (1) = {1} × V ,
y tal que es fácil calcular los primeros términos de los desarrollos de Puiseux
de las ramas de V que están sobre 0. Los métodos de continuación numéricos
calculan los primeros términos de los desarrollos de Puiseux y luego siguen
las ramas de V por el intervalo [0, 1] para obtener aproximaciones a todos los
puntos de V .
Vamos a ver cómo combinar los procedimientos numéricos basados en
homotopías poliedrales de [HS95] con nuestras técnicas de deformación —en
particular, las técnicas de deformación del Capítulo 4 (y de [BMWW04])—
21
CONTENTS
a fin de diseñar un algoritmo probabilístico simbólico para resolver sistemas
polinomiales ralos de dimensión cero con un costo cúbico en el tamaño de la
estructura combinatoria de la entrada. Para esto, en el Capítulo 5 reproduciremos los resultados de [JMSW09] —que también es de mi co-autoría.
Aproximadamente, el resultado principal de este capítulo es el siguiente
resultado (ver el Teorema 5.23 para más precisiones).
Sean f1 , . . . , fn polinomios en Q[X1 , . . . , Xn ] tal que el sistema f1 =
0, . . . , fn = 0 define una variedad afín cero-dimensional V en Cn . Sean
∆1 , . . . , ∆n ⊂ Zn≥0 los soportes de f1 , . . . , fn y supongamos que 0 ∈ ∆i
para 1 ≤ i ≤ n, y que el volumen mixto D de los polítopos de Newton
Q1 := Conv(∆1 ), . . . , Qn := Conv(∆n ) es distinto de cero.
Entonces, mediante un algoritmo probabilístico en Q, podemos calcular
la solución geométrica de V en
O(N DD′ )
operaciones
términos poli-logarítmicos), donP
P aritméticas en′ Q (omitiendo
de N := 1≤i≤n #∆i y D := 1≤i≤n M(∆, Q1 , . . . , Qi−1 , Qi+1 , . . . , Qn ); ∆
denota el símplice n-dimensional y M denota volumen mixto.
Este ítem será estudiado en detalle en el Capítulo 5.
22
CONTENTS
Chapter 1
Introduction
Given a polynomial equation system over the rationals, we want to answer
questions about its zero set. According to tradition in theoretical computer
science, problems are specified by syntactic parameters. For polynomial systems, this amounts to producing a (not necessarily uniform) algorithm that,
given integers d, n and s, for every system of n-variate polynomials over the
rationals consisting of s equations of total degree bounded by d which define
a zero-dimensional variety, computes its zero set. It can be shown that this
problem is P # -hard and one sees easily that the problem to decide wether a
given polynomial equation system over the integers is consistent is NP- and
NPC -hard ([HM93], [SS93a]).
In fact, both symbolic and numeric approaches to polynomial system
solving suffer from grave disadvantages. The numeric approach cannot be
applied directly to solve parametric, overdetermined, underdetermined, or
degenerated systems —that are of great interest (see, e.g., [Par95]). Moreover, when comparing symbolic and numeric approaches, such as the numeric
procedures of [SS93a], [SS93b], [SS93c], [SS96b], [SS94], [CS99], [BCSS98],
one can show that the latter is worse in terms of bitwise computations (see
[CHMP01] and [CJPS02]).
On the other hand, the infamous result of Mayr and Meyer ([MM82]) says
that the “Hauptproblem der Idealtheorie” (i.e., the ideal membership problem) is EXPSPACE-complete; and, in turn this is evidence that symbolic
approaches based on re-writing techniques, such as Gröbner basis computations (see [Buc85]), require exponential memory in worst case (see also,
23
24
CHAPTER 1. INTRODUCTION
e.g., [May89], [KM96] and [Küh98]). Another result, due to D. Lazard, T.
Mora, W. Masser and P. Philippon (see [Bro87]), shows that an exponential
number of computations is unavoidable when using the dense representation of polynomials. The symbolic re-writing approach is nevertheless used
by a considerable portion of the software that is currently available for this
problem.
We propose to avoid this complexity blow up of symbolic algorithms by restricting our attention to “geometric questions” or special cases of polynomial
equation systems, e.g., to those which have only finite zero sets. However,
even for the case of systems having a finite zero set, the time complexity
2
of optimal symbolic algorithms is of order sdO(n ) . This is due to the fact
that symbolic algorithms are limited to the dense (or sparse) representations
of polynomials. A way out of this dilemma is given by the representation
of polynomials by (division-free) straight-line programs which evaluate them.
But even in this case, the improvement of the worst-case complexity is limited
to sdO(n) for algorithms which are implemented using the rules of software
engineering ([HKR11], [HKR12]).
1.1.
A catalogue
The problem we undertake then is to determine classes of systems that can
be solved with “reasonable” computational resources available to scientists
and engineers these days.
Our approach grows from the findings of [GHM+ 98], [GHH+ 97],
[GHMP97], [Par95]. In these articles the authors introduce algorithms for
solving polynomial systems that rely on a homotopic deformation which is
“tracked” by means of a Newton-Hensel lifting. These procedures take as input a straight-line program computing the polynomials defining the system
under consideration and output its (geometric) solution in time polynomial
in a new parameter called the degree of the system, which is bounded by the
Bézout number of this system.
We further take on the teachings of [HKP+ 00], [GLS01], [HMW01] and
[Sch03]. In this second instalment of works, the authors isolate a deformation
algorithm based on the recently-mentioned lifting techniques, determine its
cost in terms of certain geometric parameters and produce efficient proce-
1.1. A CATALOGUE
25
dures to compute the solution of certain polynomial systems.
Before we describe these deformation techniques, we want to take a step
back to recall a result of [CGH+ 03] (cf. [HMPW98], [Par00], [GH01], [BP06]),
which shows that any robust universal algorithm solving certain polynomial
systems over the complex numbers require at least DΩ(1) arithmetic operations, where D denotes the Bézout number of the input system. For the
classes of polynomial systems in our catalogue we aim to: a) replace the
Bézout number D by parameters that reflect the geometric properties of the
input system, and b) replace the Ω(1) in the above estimate for a constant
that is as low as possible. This effort will be done not to make a dent in the
general case, as we have already explained, but in particular examples that
will constitute the items of our catalogue.
Suppose that we are given polynomials f1 , . . . , fn ∈ Q[X1 , . . . , Xn ] defining a zero-dimensional solution set V ⊂ Cn . Assume that we can define an
algebraic curve V ⊂ Cn+1 and a dominant and generically-unramified morphism π : V → C such that π −1 (1) = {1} × V holds, where V is the set of
common zeros in Cn+1 of polynomials F1 , . . . , Fn ∈ Q[T, X1 , . . . , Xn ] and the
morphism is defined by π(t, x1 , . . . , xn ) := t. Then, from a complete description of an unramified fibre π −1 (t0 ) (see Section 2.2.1 for a precise definition)
we can compute a complete description of an arbitrary fibre π −1 (t), and thus
of V . We call “lifting procedure” to an algorithm that solves this problem.
We will typically produce F1 , . . . , Fn as perturbations of f1 , . . . , fn so that
the polynomial system {F1 (t0 , X) = 0, . . . , Fn (t0 , X) = 0} is “easy” to solve
and Fi (1, X) = fi (X) holds for 1 ≤ i ≤ n. For this lifting procedure, the
polynomials F1 , . . . , Fn in Q[T, X] := Q[T, X1 , . . . , Xn ] are required to form
a regular sequence and span a radical ideal in Q[T, X]. The method requires
a point t0 ∈ Q such that its fibre under π is unramified and a linear form
U ∈ Q[X] such that U separates the points of π −1 (t0 ).
The lifting procedure computes a complete description of V from: a
straight-line program in Q[T, X] that computes F1 , . . . , Fn and a complete
description of the fibre π −1 (t0 ). The output consists of n + 1 univariate
polynomials which are given by its dense representation.
With these hypotheses, there exist #π −1 (t0 ) different n-tuples of formal
power series R := (R1 , . . . , Rn ) ∈ C[[T − t0 ]]n that are solutions of the system
F1 = 0, . . . , Fn = 0, i.e., it holds that Fi (T, R) = 0 in C[[T − t0 ]] for 1 ≤ i ≤ n.
26
The Newton-Hensel lifting we refer to computes an approximation to these
formal power series and computes a complete description of V from it.
We illustrate the Newton-Hensel lifting by depicting how the NewtonHensel operator (formally) approximates the solutions of the system under
consideration. This is by no means an algorithm that will be used later on
to solve polynomial systems.
Let ξ ∈ Cn be any point in π −1 (t0 ). Then, with our hypotheses, we
can show that there exists a unique n–tuple of formal power series R(ξ) :=
(ξ)
(ξ)
(R1 , . . . , Rn ) ∈ C[[T − t0 ]]n such that R(ξ) (0) = ξ and Fi (T, R(ξ) (T )) = 0
for every 1 ≤ i ≤ n in C[[T − t0 ]].
Write JF (X) := (∂Fi /∂Xj )1≤i,j≤n for the Jacobian matrix of F1 , . . . , Fn
with respect to X1 , . . . , Xn and let
 


X1
F1 (T, X)
 


..
NF (X) :=  ...  − (JF (T, X))−1 

.
Xn
Fn (T, X)
denote the (formal) Newton-Hensel operator. For κ ∈ Z≥0 , let
(ξ,κ)
R(ξ,κ) := (R1
, . . . , Rn(ξ,κ) ) := NFκ (T, ξ)
denote the κ–the fold iteration of the Newton-Hensel operator NF starting
at ξ. We claim that for every κ ∈ Z≥0 it holds that
det(JF (T, R(ξ,κ) )) ∈
/ (T − t0 )C[T ](T −t0 ) , and
κ
Fi (T, R(ξ,κ) ) ∈ (T − t0 )2 C[T ](T −t0 ) for 1 ≤ i ≤ n.
Moreover, it also holds that
(ξ,κ+1)
R(ξ,κ+1) agrees with R(ξ,κ) in its first 2κ powers of (T −t0 ), i.e., Ri
κ
(ξ,κ)
Ri
∈ (T − t0 )2 C[T ](T −t0 ) for 1 ≤ i ≤ n.
−
It is straightforward to show that these assertions are true when κ = 0,
because by hypotheses the ideal (F1 (t0 , X), . . . , Fn (t0 , X)) of C[X] is radical,
since the fibre π −1 (t0 ) is unramified. This implies that det(JF (t0 , ξ)) 6= 0
27
1.1. A CATALOGUE
and therefore det(JF (T, R(ξ,0) )) ∈
/ (T − t0 )C[T ](T −t0 ) . Besides, it holds that
(ξ,0)
Fi (T, R ) = Fi (T, ξ) ∈ (T − t0 )C[T ](T −t0 ) for 1 ≤ i ≤ n. From the identity
R(ξ,1) = R(ξ,0) + JF (T, R(ξ,0) )−1 F (T, R(ξ,0) )
(ξ,1)
we deduce that Ri
(ξ,0)
− Ri
∈ (T − t0 )C[T ](T −t0 ) for 1 ≤ i ≤ n.
The inductive step follows from the definition of the operator NF and the
Taylor expansions of the Fi and det(JF ) as power series in (T − t0 ). For each
1 ≤ i ≤ n, we multiply each member of the equality


F1 (T, R(ξ,κ) )


..
R(ξ,κ+1) − R(ξ,κ) = −JF−1 (T, R(ξ,κ) ) · 

.
Fn (T, R(ξ,κ) )
by the i–th row of JF (T, R(ξ,κ) ) and deduce that
JF (T, R(ξ,κ) )i · (R(ξ,κ+1) − R(ξ,κ) ) = −Fi (T, R(ξ,κ) ).
(1.1)
Combining (1.1) with the congruence relation
Fi (T, R
(ξ,κ+1)
) ≡ Fi (T, R
(ξ,κ)
n
X
∂Fi
(ξ,κ+1)
(ξ,κ)
)+
(T, R(ξ,κ) ) · (Rj
− Rj )
∂Xj
j=1
mod (R(ξ,κ+1) − R(ξ,κ) )2 in C[[T − t0 ]](T −t0 ) , we deduce that Fi (T, R(ξ,κ+1) ) ≡ 0
mod (R(ξ,κ+1) − R(ξ,κ) )2 in C[[T − t0 ]](T −t0 ) . From the inductive hypotheses we
κ
κ+1
obtain that (R(ξ,κ+1) − R(ξ,κ) )2 ⊂ ((T − t0 )2 )2 ⊂ (T − t0 )2
which proves
the second claim in the inductive step.
The last claim in the inductive step follows from the equality


F1 (T, R(ξ,κ+1) )


..
R(ξ,κ+2) − R(ξ,κ+1) = −JF−1 (T, R(ξ,κ+1) ) · 

.
Fn (T, R(ξ,κ+1) )
and what we just proved. On the other hand, from the Taylor expansion of
the Jacobian determinant det(JF ) we infer that
det(JF (T, R(ξ,κ+1) ) ≡ det(JF (T, R(ξ,κ) ) +
n
X
∂ det(JF )
(ξ,κ+1)
(ξ,κ)
(T, R(ξ,κ) ) · (Rj
− Rj )
∂X
j
j=1
28
mod (R(ξ,κ+1) −R(ξ,κ) )2 in C[[T −t0 ]](T −t0 ) . From the first claim in the inductive
hypotheses we deduce its counterpart in the inductive step.
We shall produce different variations of lifting procedures which follow
steps that cannot be inferred in a straightforward fashion from the claims
above. These rely mainly on the algorithmic contributions of [GLS01] and
[Sch03] and use roughly O(deg V deg π) arithmetic operations in Q, where
deg V and deg π denote the degree of the variety V and the degree of the morphism π respectively (see Section 2.3 for definitions of complexity measures
and Section 2.4 for a precise statement). In our setting, deg π ≤ deg V ≤ D
holds, where D is the Bézout number of the system defining V.
With this algorithm available, a critical point for the application of this
homotopy-lifting method is to obtain morphisms π as above of “low degree”
with a fibre that is “easy to solve”.
1.1.1.
Generalized Pham Systems
Our catalogue starts with a significant class of zero–dimensional square
polynomial systems with rational coefficients over the complex numbers,
called generalized Pham systems (see [PS04] and [DMW09]) or strict complete
intersections (see [CDS96]), which arise in connection with several problems
in computational algebraic geometry (see, e.g., [MT00], [HM00]). A generalized Pham system may be roughly described as the result of a deformation
of an isolated projective complete–intersection singularity and corresponds
to the intuitive notion of a system with “no points at infinity” (see [PS04,
Remark 17] or [CDS96, Section 1]).
An n–dimensional generalized Pham system is defined by n polynomials
of the form
φ1 + ϕ1 , . . . , φ n + ϕn ,
where for each 1 ≤ i ≤ n, it holds that φi ∈ Q[X1 , . . . , Xn ] is homogeneous
and ϕi is a polynomial (not necessarily homogenous) of degree less than
di := deg φi , and such that φ1 , . . . , φn define the empty projective variety of
Pn−1 (C).
A particular example of a generalized Pham system is an (n–dimensional)
29
1.1. A CATALOGUE
Pham system, defined by n polynomials of the form
X1d1 + ϕ1 , . . . , Xndn + ϕn ,
where ϕi ∈ Q[X1 , . . . , Xn ] has degree less than di for 1 ≤ i ≤ n (in Section 3.1
we shall exhibit other useful examples of generalized Pham systems).
The solutions of a Pham system f1 = 0, . . . , fn = 0 in Q[X1 , . . . , Xn ]
can be computed with an adaptation of [Sch03] or even using a result in
[BMWW04, Section 5], a paper that I co-authored. One would apply the
“projection algorithm” (as in the cited articles) to the deformation πW : W →
A1 determined by the morphism πW (t, x) := t, where the variety W ⊂ An+1
consists in the common zeros of the system
ai Xidi + T (fi − ai Xidi ) + bi (1 − T ) = 0 (1 ≤ i ≤ n),
where ai stands for the coefficient of X di in fi and bi is a randomly chosen
rational for 1 ≤ i ≤ n. In [BMWW04] we exhibited a procedure solving
families of Pham systems, of quadratic complexity, using the homotopy-lifting
procedure in a straightforward fashion (in the case where d1 = . . . = dn ).
This procedure is described later in Chapter 4 under a different context.
Unfortunately, the coordinate ring of a generalized Pham system lacks the
simple monomial structure arising in a Pham system and therefore this procedure (cf. [BMWW04]) —but also the procedures of [MP97], [GLGV98],
[MP00], [MT00]— leads to an algorithm with more than quadratic complexity in the Bézout number of the input system. Explicitly, considering
the product deg(πW ) deg W, which is the dominant term of the complexity of this algorithm, in this case we have deg(πW ) = D := d1 · · · dn and
deg W ≤ E := (d1 + 1) · · · (dn + 1) according to the Bézout inequality. In
particular, for n ≫ max{d1 , . . . , dn } we have E ≫ D and hence DE ≫ D2 .
Notwithstanding, we shall show that it is still possible to produce a probabilistic algorithm which solves generalized Pham systems with quadratic
complexity in the Bézout number of the input system. The algorithm will
underly not one, but a sequence of n + 1 homotopies in artificially produced
curves that will lead to the computation of the solution of the original system.
The design of this algorithm and the proof of the required mathematical
facts are the subject of Chapter 3. This is one of the two main results
covered in [DMW09], which I co-authored. The second result states that
30
a robust universal algorithm solving generalized Pham systems has (worst–
case) complexity of order DΩ(1) , where D is the Bézout number of the input
system. This bound is independent of the representation of input and output,
but the value of the exponent underlying the Ω–notation does depend on such
a representation. For example, if the usual dense or sparse representation (the
list of all or of all nonzero coefficients) is used, then the complexity of the
corresponding algorithm is of order Ω(D), while for the straight–line program
representation a lower bound of order Ω(D1/2 ) is achieved. This shows that
our procedure is almost optimal (in its class).
1.1.2.
Lifting from ramified fibres
The second item in our catalogue serves as a generalization for the algorithm we presented and used for the previous item. Let V ⊂ Cn+1 be a
Q–definable algebraic curve, and let us assume that the morphism π : V → C
induced by the canonical projection in the first coordinate is dominant and
generically unramified. Suppose that we are given the infinitesimal structure of a finite and ramified fibre π −1 (tr ). We want to compute a complete
description of an arbitrary fibre π −1 (t).
While in the case of an unramified fibre, we could use a Newton-Hensel
lifting to approximate the solutions and derive a complete description of an
arbitrary fibre π −1 (t), this is no longer possible when the fibre is ramified.
Suppose that the fibre π −1 (tu ) is unramified for a given tu ∈ Q. Recall that
the quotient field C((T − tu )) of the ring of formal power series C[[T − tu ]] is
not algebraically closed, and that the field of Puiseux series Q(T − tu )∗ is an
algebraic closure of Q((T − tu )) (see, e.g., [Wal50]). Then, all the solutions in
An (C(T − tu )∗ ) of the system
F1 (T, X) = 0, . . . , Fn (T, X) = 0
(1.2)
belong to C[[T −tu ]]n as we have already shown in Section 1.1. However, when
(1.2) is considered as a system in An (C(T − tr )∗ ), the argument we used in
Section 1.1 fails; in fact, not all its solutions necessarily belong to C[[T − tr ]]n .
This implies that we can not use our lifting algorithm in this context.
Our way out of this dilemma is by requiring a complete description of
the set of singular parts of the Puiseux expansions of the branches of V lying
above tr (see Section 4.1 for further details). In Chapter 4 we will exhibit an
1.1. A CATALOGUE
31
algorithm which computes a complete description of an arbitrary fibre π −1 (t)
and which requires roughly O(deg V(deg π)α ) operations in Q, where α = 1
in several important cases. This is the main result of [BMWW04], which I
co-authored. This result extends and improves the procedures in [HKP+ 00]
and [Sch03] (cf. [DMW09]).
We shall also provide examples where the families underly some “easy
fibres” that are ramified. In such cases, this new lifting technique allows us
to compute a complete description of any member of the family V.
For example, for n, d ∈ N, consider the Pham system
f1 := X1d − ϕ1 (X), . . . , fn := Xnd − ϕn (X),
the curve defined by the system
F1 := X1d + T ϕ1 (X), . . . , Fn := Xnd + T ϕn (X),
and the morphism π induced by the projection on the T coordinate. Then,
the fibre π −1 (0) = {0} has only one point and is ramified. In this case, the
singular parts of the branches of the curve {F1 = 0, . . . , Fn = 0} lying above
T = 0 are given by
1/d
(ξ i1 α1 T 1/d , . . . , ξ in αn1/d T 1/d )
for 0 ≤ i1 , . . . , in ≤ d − 1, where α := (α1 , . . . , αn ) := (ϕ1 (0), . . . , ϕn (0)). As
we will show, one may apply our lifting techniques for ramified fibres in order
to solve this example with roughly O(T deg V deg π) = O(Td2n ) arithmetic
operations in Q.
The family of polynomial systems below shall serve as another example
where this new algorithm can be applied.
1.1.3.
Sparse systems and polyhedral homotopies
The catalogue concludes with a procedure for computing all solutions
to zero-dimensional sparse polynomial systems (where the set of non-zero
coefficients of the defining polynomials is “small”) and which uses a polyhedral
homotopy.
The infamous results by D.N. Bernstein, A.G. Kushnirenko and A.G.
Khovanski ([Ber75], [Kus76], [Kho78]) bound the number of solutions of a
32
polynomial system in terms of a combinatorial invariant associated to the
set of exponents of the monomials arising with nonzero coefficients in the
defining polynomials. More precisely, the Bernstein-Kushnirenko-Khovanski
(BKK for short) theorem asserts that the number of isolated solutions in the
n-dimensional complex torus (C∗ )n of a polynomial system of n equations
in n unknowns is bounded by the mixed volume of the family of Newton
polytopes of the corresponding polynomials.
Numeric (homotopy continuation) methods for sparse systems are typically based on a specific family of deformations called polyhedral homotopies
([HS95], [VVC94], [VGC96], [Roj03]). Polyhedral homotopies preserve the
Newton polytope of the input polynomials and yield an effective version of
the BKK theorem (see, e.g., [HS95], [HS97]).
Suppose that we are given a zero-dimensional (∆1 , . . . , ∆n )-sparse system
defined by n polynomials f1 , . . . , fn ∈ Q[X1 , . . . , Xn ], where ∆1 , . . . , ∆n ⊂
Zn are the supports of f1 , . . . , fn . Let V ⊂ (C∗ )n be the variety defined
by the common zeros of f1 , . . . , fn in (C∗ )n . Then a polyhedral homotopy
consists in an algebraic curve V ⊂ (C∗ )n+1 such that the projection π : V →
C∗ , π(t, x) := t onto the first coordinate is dominant with generically finite
fibres whose degree is the mixed volume M V (conv(∆1 ), . . . , conv(∆n )) of the
convex hulls of ∆1 , . . . , ∆n , the identity π −1 (1) = {1} × V holds and the first
terms of the Puiseux expansions of the branches of V lying above 0 can be
easily computed. Numerical continuation methods compute the first terms
of these Puiseux expansions and then follow the branches of V along the
interval [0, 1] to obtain approximations to all the points of the input variety
V.
We will show how to combine the homotopic procedures of [HS95] with
our deformation techniques, particularly in the version of Chapter 4 (and
[BMWW04]), in order to derive a symbolic probabilistic algorithm for solving
sparse zero-dimensional polynomial systems with cubic cost in the size of the
combinatorial structure of the input system. This chapter reproduces the
results of [JMSW09] —an article that I co-authored.
Our main result may be stated as follows (see Theorem 5.23 for a precise
statement):
Let f1 , . . . , fn be polynomials in Q[X1 , . . . , Xn ] such that the system
f1 = 0, . . . , fn = 0 defines a zero-dimensional affine sub-variety V of Cn .
Denote by ∆1 , . . . , ∆n ⊂ Zn≥0 the supports of f1 , . . . , fn . Assume that 0 ∈ ∆i
1.1. A CATALOGUE
33
for 1 ≤ i ≤ n and the mixed volume D of the Newton polytopes Q1 :=
Conv(∆1 ), . . . , Qn := Conv(∆n ) is nonzero.
Then, we can probabilistically compute a geometric solution of the vari′
ety V using O(N DD
P in Q (omitting poly-logarithmic
P ) arithmetic operations
terms), with N := 1≤i≤n #∆i and D′ := 1≤i≤n M(∆, Q1 , . . . , Qi−1 , Qi+1 , . . . , Qn ),
where ∆ denotes the standard n-dimensional simplex and M stands for
mixed volume.
All this shall be covered in Chapter 5.
34
Chapter 2
Preliminaries
2.1.
Notation
Let K be a zero-characteristic field and let K be an algebraically closed
field containing K. We shall use the standard notation for the set of integers
Z, positive integers N, and the fields of rationals Q, algebraic numbers Q and
complex numbers C.
For any positive real number a, by log a we denote its binary logarithm
(in base 2).
Let n ∈ N. Let X1 , . . . , Xn be indeterminates over K. Unless otherwise stated X shall stand for the vector X := (X1 , . . . , Xn ). We denote by
K[X1 , . . . , Xn ] (or K[X] for short) the ring of polynomials in n variables over
K. We shall also consider the elements in K[X] as functions from K n to K.
2.2.
Geometry
n
Let An := An K be the n–dimensional affine space K .
Let be given a set of polynomials S ⊆ K[X1 , . . . , Xn ]. The set of common
zeroes of the polynomials in S is denoted by
V (S) := {x ∈ An : f (x) = 0 ∀f ∈ S}.
35
36
CHAPTER 2. PRELIMINARIES
We say that V ⊂ An is a K–definable algebraic variety if and only if there
exists a set S ⊆ K[X1 , . . . , Xn ] such that V = V (S). Unless specifically
noted, variety shall mean K–definable algebraic variety.
In the affine space An we define the Zariski topology over K,where the
closed sets are all the K–definable algebraic varieties V ⊂ An K .
In particular we shall consider fields K such as the fields of rational numbers Q, complex numbers C, rational functions with rational coefficients
Q(T ) or its algebraic closure Q(T ) = Q(T )∗ , namely the field of Puiseux
series, where T is a new indeterminate. Further, we shall consider the rings
of n–variate polynomials over these fields.
An algebraic variety V is irreducible if it is not the union of two varieties
which are proper subsets of V . Note that since An is Noetherian, every
algebraic variety V is the union of a finite number of irreducible algebraic
varieties. Moreover, the irreducible varieties in such a decomposition are
unique up to reordering and are called the irreducible components of V .
The dimension of a variety V of An , which we denote by dimAn V or dim V ,
is the supremum over the integers r such that
W0 ⊂ . . . ⊂ Wr
is a strictly ascending sequence of irreducible sub-varieties of V —and it
is the same as the Krull dimension of the ring K[X1 , . . . , Xn ]/I(V ). Here
I(V ) := {f ∈ K[X] : f (x) = 0 ∀x ∈ V } is the ideal defined by all the
polynomials that vanish on V . A variety V of An is said equidimensional if
all its irreducible components have the same dimension.
Let W and V be sub-varieties of An such that W ⊆ V and V is irreducible.
When W is irreducible, the co-dimension of W in V is the supremum over
the integers r such that W ⊂ W1 ⊂ . . . ⊂ Wr is a strictly increasing chain of
irreducible varieties contained in V . For an arbitrary sub-variety W of V , its
co-dimension is defined as the infimum of the co-dimensions of its irreducible
components. An hypersurface of V is an equidimensional sub-variety of codimension 1. When V = An , and in other important cases, a variety V
is an hypersurface if and only if it is the set of zeros of one non-constant
polynomial f ∈ K[X1 , . . . , Xn ].
Let V ⊂ An be a nonempty irreducible variety of dimension dim V = r.
37
2.2. GEOMETRY
The degree deg V of V is defined as
deg V := degAn V := sup # (H1 ∩ . . . ∩ Hr ∩ V ) : H1 , . . . , Hr are
affine hyperplanes of A , and the intersection is finite .
n
This is always a finite number. If V is an arbitrary algebraic sub-variety
of An , following [Hei83] and [Ful84], we define its degree as the sum of the
degrees of all its irreducible components. If V and W are subvarieties of An ,
then the Bézout inequality ([Hei83]; see also [Ful84], [Vog84]) asserts that:
deg(V ∩ W ) ≤ deg V deg W.
(2.1)
Let V ⊆ An and W ⊆ Am be K–definable varieties. The restriction of a
polynomial function P ∈ K[X] to V is called a regular function of V . The
set of regular functions of V is a ring that we denote by K[V ]. A morphism
of varieties is a map ϕ from V to W given by regular functions ϕ1 , . . . , ϕm ∈
K[V ] such that for every t ∈ V it holds that ϕ(t) = (ϕ1 (t), . . . , ϕm (t)) ∈ W .
The total quotient ring of V , denoted by K(V ), is the ring resulting from
the localisation S −1 · K[V ], where S is the set of nonzero divisors of K[V ].
The members of K(V ) are called rational functions. We note that if V is
irreducible, then K(V ) is the field of fractions of K[VQ
]; however, in the
general case it is the product of the fields of fractions
K(Ci ), where Ci
runs over the irreducible components of V .
A morphism (of varieties) ϕ : V → W is said dominant if its image ϕ(V )
is dense in W (in the Zariski topology).
A morphism of varieties ϕ : V → W induces a morphism of rings
ϕ : K[W ] → K[V ] from K[W ] = K[Y1 , . . . , Ym ]/I(W ) to K[V ] = K[X1 , . . . ,
Xn ]/I(V ), which is univocally determined by mapping the regular functions
of K[W ] defined by the variables Yi , for 1 ≤ i ≤ m, to ϕ∗ (Yi ) = ϕi respectively.
∗
A morphism ϕ : V → W is said a finite morphism if K[V ] is integral over
K[W ]; in particular, its fibres are nonempty and finite (i.e., 0 < #ϕ−1 (y) <
∞ for every y ∈ ϕ(V )).
38
Let V and W be equidimensional sub-varieties of An and Am respectively,
and of equal dimension dim V = dim W . Let ϕ : V → W be a dominant
morphism. When V is irreducible, we define the degree of ϕ as
deg ϕ := [K(V ) : ϕ∗ (K(W ))],
where the brackets denote the degree of the field extension ϕ∗ (K(W )) →
K(V ). More generally, if the decomposition of V into irreducible components
is given by V = ∪Ci and each of the restrictions ϕ|Ci : Ci → W is dominant,
we define the degree of ϕ as the sum of the degrees of the restrictions ϕ|Ci ,
namely
X
X
deg ϕ :=
deg(ϕ|Ci ) =
[K(Ci ) : ϕ∗ (K(W ))].
i
i
n+1
In particular, when V ⊂ A
is equidimensional of dimension 1, V = C1 ∪
· · · ∪ CN is a decomposition into irreducible components, and the morphism
π : V → A1 defined by π(x1 , . . . , xn+1 ) := x1 is such that the restriction
π|
PCNi is dominant for 1 ≤ i ≤ N , then the degree of π is the integer D =
i=1 [K(Ci ) : K(X1 )], where [K(Ci ) : K(X1 )] denotes the degree of the
(finite) field extension K(X1 ) → K(Ci ) for 1 ≤ i ≤ N .
Polynomials f1 , . . . , fs are said to form a regular sequence in K[X1 , . . . , Xn ]
if f1 is not zero and fi is not a zero divisor in K[X1 , . . . , Xn ]/(f1 , . . . , fi−1 ) for
i = 2, . . . , s. In such a case, the affine variety V := V (f1 , . . . , fs ) ⊂ An that
they define is equidimensional of dimension n − s, and is called a set-theoretic
complete intersection variety. Further, if the ideal (f1 , . . . , fs ) generated by
f1 , . . . , fs is radical, then we say that V is ideal-theoretic complete intersection.
Let V be an equidimensional sub-variety of An of dimension dim V = r defined by polynomials f1 , . . . , fn−r , and such that the ideal I := (f1 , . . . , fn−r )
of K[X1 , . . . , Xn ] is radical. Then Noether’s normalization theorem states that
there exist linear forms Y1 , . . . , Yr ∈ K[X1 , . . . , Xn ]/I algebraically independent over K[X1 , . . . , Xn ]/I such that the extension
K[Y1 , . . . , Yr ] → K[X1 , . . . , Xn ]/I
is finite. In such a case, we say that Y1 , . . . , Yr are in Noether position with
respect to the variety V (f1 , . . . , fn−r ).
Let f ∈ K[X] be a regular function of An . Then the differential of f at
a point x ∈ An is the linear map
dx f : K n → K
Pn
ξ 7→
i=1
∂f
(x)ξi
∂Xi
.
39
2.2. GEOMETRY
Let V ⊂ An be a variety and x ∈ V . We define the tangent space of V at x
by
Tx V = {ξ ∈ An : dx g(ξ) = 0 ∀g ∈ I(V )}.
Let W ⊂ Am be an algebraic variety. Let us assume that ϕ : V → W
is a finite morphism. Let y = (y1 , . . . , ym ) ∈ W and let My be the maximal ideal of K[Y1 , . . . , Ym ] generated by the polynomials Y1 − y1 , . . . , Ym −
ym . We borrow two notions from the realm of schemes. We say that the
K[W ]My /My K[W ]My –algebra
K[V ]My /My K[V ]My
represents the fibre of y under ϕ.
We say that ϕ is unramified at a point x ∈ V if, and only, if the differential
map dx ϕ : Tx V → Tϕ(x) W is injective. We say that the fibre of y ∈ ϕ(V ) ⊂ W
is unramified if ϕ is unramified at every x ∈ ϕ−1 (y).
Assume further that V and W are equidimensional and of equal dimension
r := dim V = dim W . Then, unramifiedness for the fibre of y = ϕ(x) under
ϕ is equivalent to K[V ]My /My K[V ]My being a product of fields. In fact, it
must hold that
deg ϕ
K[V ]My /My K[V ]My = K[W ]My /My K[W ]My
.
A particular case of this situation which will be important for us is the
following: assume that V := V (f1 , . . . , fn−r ) is an ideal-theoretic complete
intersection defined by polynomials f1 , . . . , fn−r ∈ K[X1 , . . . , Xn ], the variables X1 , . . . , Xr are in Noether position and let ϕ : V → Ar be the linear
mapping defined by ϕ(x) := (x1 , . . . , xr ). Then ϕ is unramified at a point
y ∈ Ar if and only if
!
∂fi
(x) 6= 0
det
∂Xj+r 1≤i,j≤n−r
at every x ∈ π −1 (y).
With the assumptions above, we say that ϕ verifies a generic condition,
e.g., flatness or unramifiedness, if this condition holds for the generic fibre
K(X1 , . . . , Xr )[Xr+1 , . . . , Xn ]/(f1 , . . . , fn−r ). In particular, ϕ is generically
40
unramified if and only if the Jacobian determinant
!
∂fi
det
∂Xr+j 1≤i,j≤n−r
is a unit in K(X1 , . . . , Xr )[Xr+1 , . . . , Xn ]/(f1 , . . . , fn−r ).
2.2.1.
Geometric Solutions
The notion of a geometric solution of an algebraic variety was first introduced in the works of Kronecker and König in the last years of the XIXth
century. Nowadays, geometric solutions are widely used in computer algebra
to represent algebraic varieties, especially in the zero-dimensional case.
We first introduce this notion in the case of zero-dimensional varieties
and then extend it to curves. Let V = {ξ (1), . . . , ξ (D)} be a zero-dimensional
sub-variety of An (K) defined over K. Let Y be a new indeterminate. A
geometric solution of V consists of
a linear form u = u1 X1 +· · ·+un Xn ∈ K[X] which separates the points
of V , i.e., satisfying u(ξ (i) ) 6= u(ξ (k) ) for i 6= k,
Q
the minimal polynomial mu := 1≤i≤D (Y − u(ξ (i) )) ∈ K[Y ] of u in V ,
polynomials w1 , . . . , wn ∈ K[Y ], with deg wj < D for every 1 ≤ j ≤ n,
satisfying
n
V = {(w1 (η), . . . , wn (η)) ∈ K : η ∈ K, mu (η) = 0}.
The linear form u is also a primitive element of the ring extension K →
K[V ] (or of V ).
In the sequel, we shall be given a polynomial system f1 = 0, . . . , fn = 0
of n-variate polynomials of Q[X] defining a zero-dimensional affine variety
V ⊂ An (C). We shall consider the system f1 = 0, . . . , fn = 0 (symbolically)
“solved” if we obtain a geometric solution of V as defined above.
Example. Let f1 , f2 ∈ Q[X1 , X2 ] be the following polynomials:
f1 := X13 − 3X12 X2 + 3X1 X22 − X23 − 11X1 + 9X2 + 7,
f2 := X12 − 2X1 X2 + X22 − 3X1 + 2X2 + 1,
2.2. GEOMETRY
41
which define the zero-dimensional variety V := {(4, 1), (0, −1), (9, 11)} in C2 .
Let u := X1 − X2 ∈ Q[X1 , X2 ]. Note that u is a separating linear form for
V . The geometric solution of V associated with u consists of
the minimal polynomial mu := (Y − 3)(Y − 1)(Y + 2) = Y 3 − 2Y 2 −
5Y + 6,
the polynomials w1 := Y 2 − 2Y + 1 and w2 := Y 2 − 3Y + 1, which
satisfy the identities (w1 (3), w2 (3)) = (4,1), (w1 (1), w2 (1)) = (0,−1) and
(w1 (−2), w2 (−2)) = (9, 11).
The notion of geometric solution can be extended to equidimensional
varieties of positive dimension. For our purposes, it will suffice to consider
the case of curves.
Suppose that we are given a curve V ⊂ An+1 defined by polynomials
f1 , . . . , fn ∈ K[T, X] = K[T, X1 , . . . , Xn ]. Assume that for each irreducible
component C of V , the identity I(C) ∩ K[T ] = {0} holds. This implies
that the morphism π|C : C → A1 defined by π(t, x) := t is dominant for
each irreducible component C of V . Let u be a nonzero linear form of K[X]
such that u separates the points of a generic fibre π −1 (t) (such a linear form
will be called a primitive element of the ring extension K(T ) → K(V ) or
of V ). Let πu : V → A2 be the morphism defined by πu (x, t) := (t, u(x)).
Our assumptions on V imply that the Zariski closure πu (V ) of the image of
V under πu is an hypersurface of A2 defined over K. Then there exists a
unique (up to scaling by nonzero elements of K) polynomial Mu ∈ K[T, Y ]
of minimal degree defining πu (V ). Let mu ∈ K(T )[Y ] denote the (unique)
monic multiple of Mu with degY (mu ) = degY (Mu ). We call mu the minimal
polynomial of u in V . In these terms, a geometric solution of the curve V
consists of
the linear form u ∈ K[X],
the minimal polynomial mu ∈ K(T )[Y ],
u
Xi = vi in K(T ) ⊗ K[V ]
elements v1 , . . . , vn of K(T )[Y ] such that ∂m
∂Y
and degY (vi ) < degY (mu ) holds for 1 ≤ i ≤ n.
42
2.3.
Complexity
Algorithms in computational algebraic geometry are usually described
using the standard dense (or sparse) complexity model, i.e., encoding multivariate polynomials by means of the vector of all (or of all nonzero) coefficients. Taking
into account that a generic n–variate polynomial of degree d
d+n
has n = O(dn ) nonzero coefficients, we see that the dense representation
of multivariate polynomials requires an exponential size, and their manipulation usually requires an exponential number of arithmetic operations with
respect to the parameters d and n. In order to avoid this exponential behavior, we are going to use an alternative encoding of input and intermediate
results of our computations by means of straight-line programs (cf. [BCS97]).
2.3.1.
Complexity model
A straight-line program β in Q(X) := Q(X1 , . . . , Xn ) is a finite sequence of
rational functions (f1 , . . . , fk ) ∈ Q(X)k such that for 1 ≤ i ≤ k, the function
fi is an element of the set {X1 , . . . , Xn }, or an element of Q (a parameter),
or there exist 1 ≤ i1 , i2 < i such that fi = fi1 ◦i fi2 holds, where ◦i is one of
the arithmetic operations +, −, ×, ÷. The straight-line program β is called
division–free if ◦i is different from ÷ for 1 ≤ i ≤ k. A natural measure of the
complexity of β is its time or length (cf. [BCS97]), which is the total number
of arithmetic operations performed during the evaluation process defined by
β. We say that the straight-line program β computes or represents a subset S
of Q(X) if S ⊂ {f1 , . . . , fk } holds.
Our model of computation is based on the concept of straight-line programs. However, a model of computation consisting only of straight-line
programs is not expressive enough for our purposes. Therefore we allow our
model to include decisions and selections (subject to previous decisions). For
this reason we shall also consider computation trees, which are straight-line
programs with branchings. The evaluation time of a given computation tree
is defined similarly to the case of straight-line programs as the maximum
evaluation time over the different branches (see, e.g., [vzG86], [BCS97] for
more details on the notion of computation trees).
In the future, when we refer to algorithms we shall mean computation
trees, and when we refer to their complexity or evaluation time we shall mean
2.3. COMPLEXITY
43
the number of arithmetic operations in Q they perform.
Our algorithms are of Monte Carlo or BPP type (see, e.g., [BDG88],
[Zip93], [vzGG99]), i.e., they return the correct output with probability at
least a fixed value strictly greater than 1/2. This means that the error
probability can be made arbitrarily small by iteration of the algorithms. On
the other hand, our algorithms do not seem to be of Las Vegas or ZPP type,
i.e., we have no means of checking the correctness of our output results.
We observe that the probabilistic aspect of our algorithms is related to the
random choice of points outside certain Zariski closed subsets of suitable
affine spaces, whose probability of success is explicitly estimated.
2.3.2.
Probabilistic identity testing
A difficult problem to handle efficiently in the manipulation of multivariate polynomials given by straight–line programs is the so-called identity
testing problem: given a straight-line program over K representing two polynomials f and g of K[X] := K[X1 , . . . , Xn ], decide whether f and g represent
n
the same polynomial function on K . Indeed, all known deterministic algorithms solving this problem have complexity at least (max{deg f, deg g})Ω(n) .
We are going to use probabilistic algorithms to solve the identity testing problem, based on the following result, know as the Zippel-Schwartz Theorem.
Theorem 2.1 ([vzGG99, Lemma 6.44]). Let f be a nonzero polynomial of
C[X] of degree at most d and let S be a finite subset of C. Then the number
of zeros of f in S n is at most d(#S)n−1 .
For the analysis of our algorithms, we shall interpret the statement of
Theorem 2.1 in terms of probabilities. More precisely, given a nonzero polynomial f in C[X] of degree at most d, we conclude from Theorem 2.1 that
the probability of picking a random point a in S n such that f (a) = 0 holds
is bounded from above by d/#S (assuming a uniform distribution of probability on the elements of S n ).
2.3.3.
Basic complexity estimates
In order to estimate the complexity of our procedures we shall frequently
use the notation M(m) := m log2 m log log m. Recall that log denotes the
44
binary logarithm. Let R be a commutative ring of characteristic zero with
unity. We recall that the number of arithmetic operations in R necessary
to compute the multiplication or division with remainder of two univariate
polynomials in R[T ] of degree at most m is O M(m)/ log(m) (cf. [vzGG99],
[BP94]). Multipoint evaluation and interpolation of univariate polynomials
of R[T ] of degree
m at invertible points a1 , . . . , am ∈ R can be performed
with O M(m) arithmetic operations in R (see, e.g., [BLS03]).
If R = K is a field, then we shall use algorithms based on the Extended Euclidean Algorithm (EEA for short) in order to compute the gcd
or resultant of two univariate polynomials in K[T ] of degree at most m
with O M(m) arithmetic operations in K (cf. [vzGG99], [BP94]). We use
Padé approximation in order to compute the dense representation of the
numerator and denominator of a rational function f = p/q ∈ K(T ) with
max{deg p, deg q} ≤ m from its Taylor series expansion up to order 2m.
This also requires O(M(m)) arithmetic operations in K ([vzGG99], [BP94]).
For brevity, we will denote by Ω the exponent that appears in the complexity estimate O(nΩ ) for the multiplication of two (n × n)-matrices with
coefficients in Q. We remark that the (theoretical) bound Ω < 2.376 is
typically impractical and we prefer to take Ω := log 7 ∼ 2.81 (cf. [BP94]).
Finally, sometimes we will refer to rough complexity estimates and omit
poly-logarithmic terms from the estimate. For example, a rough estimate for
O(m log2 m log log m) may be O(m).
2.4.
Deformation algorithms based on lifting
techniques
We discuss in depth the lifting technique mentioned earlier in Section 1.1.
This technique relies on a formal version of the Newton-Hensel lifting algorithm that was introduced, in the context of polynomial equation solving, in
[GHM+ 98] and [GHH+ 97]. Its essentials became described in [HKP+ 00] and
its study was continued in [BMWW04], [Sch03], and [JMSW09].
Deformation algorithms based on lifting techniques appear in flavours of
the following statement:
2.4. DEFORMATION ALGORITHMS
45
Let n be a positive integer and let F1 , . . . , Fn ∈ Q[T, X1 , . . . , Xn ] be
polynomials defining an equidimensional algebraic variety V ⊆ An+1 of dimension 1. Let π : V → A1 be the morphism defined by π(t, x) := t for
(t, x) := (t, x1 , . . . , xn ) ∈ V. Assume π to be finite and generically unramified. Let g ∈ Q[V] be the element in the coordinate ring of V that is
defined by a polynomial G ∈ Q[T, X]. Then, there exists a unique polynomial P ∈ Q[T, Y ] which is the (minimal) integral dependence equation that
is satisfied by g in the extension Q[T ] → Q[V].
Suppose furthermore that we are given a point t0 ∈ Q such that its fibre
under π is unramified and that G maps π −1 (t0 ) onto degY P distinct points.
The projection problem is that of computing P given the polynomials
F1 , . . . , Fn , the polynomial G and a geometric solution of the affine zerodimensional variety π −1 (t0 ). We call P the projection of g onto V.
The main algorithm in [HKP+ 00] solves the projection problem when
F1 , . . . , Fn , G are given by a straight-line program in Q[T, X] such that F1 ,
. . . , Fn form a regular sequence and span a radical ideal in Q[T, X], and the
geometric solution of the fibre π −1 (t0 ) is provided by the dense representation
of the univariate polynomials in Q[Y ] that define it. The output P is provided
in its dense representation. The computation takes roughly
O((T + n4 )(deg π)3 degT P ) = O(T n4 (deg π)3 deg V degX G)
(2.2)
arithmetic operations in Q (omitting poly-logarithmic terms). Here, degT P
denotes the partial degree of P in the variable T and degX G denotes the
partial degree of G in the variables X1 , . . . , Xn , deg V stands for the degree
of the affine variety V and deg π stands for the degree of the morphism π.
Note that when G is a generic linear form (i.e., degX G = 1), its projection
is actually the minimal polynomial in a geometric solution (that has G as
its separating linear form). In this case, degY P = deg π since G separates
points on π −1 (t0 ), so #π −1 (t0 ) = deg π, and therefore P (t0 , Y ) has deg π
distinct roots.
It can also be shown (see, e.g., [DMW09]) that a complexity estimate of
order deg π O(1) deg V O(1) is unavoidable; in fact, by tracing how these parameters are introduced one sees that deg V is the accuracy we require for our
approximations and deg π is the number of approximations we should compute (i.e., one per point in the zero-dimensional variety π −1 (t0 )). However,
46
a reduction to roughly
O(n5 T deg π(deg V) degX G)
arithmetic operations in Q is possible (see Theorem 2.2 below; see also
[GLS01] and [Sch03]).
2.4.1.
The lifting process
Let the notions and notations from the preceding section hold. Let us
see how to compute the minimal polynomial P of the projection of V defined
by a linear form U ∈ K[X] (i.e., the minimal integral dependence equation
in K[T ] → K[V] of a representative of U ∈ K[X] in K[V]). Extending
this procedure to an arbitrary polynomial G(T, X) is straightforward but
unnecessary for what follows (see [HKP+ 00] for details).
We discuss the extension of [Sch03] to a method of [GLS01] for computing
a geometric solution of the curve V .
Theorem 2.2. Let F1 , . . . , Fn ∈ K[T, X], t0 be a point in K, U be a linear
form in K[X1 , . . . , Xn ] and q, w1 , . . . , wn ∈ K[Y ]. Let V ⊂ An+1 be the
variety defined by F1 , . . . , Fn , and let π : V → A1 be the morphism defined by
the projection π(t, x) := t. Assume that:
The polynomials F1 , . . . , Fn form a regular sequence and span a radical
ideal in K[T, X]; also, they are computed by a straight-line program β
in T operations over K.
The morphism π is dominant and generically unramified.
It holds that #π −1 (t0 ) = deg π.
The coordinate functions defined by the linear form U are primitive
elements in both the extension K → K[Vt0 ] and K[T ] → K[V], where
Vt0 is the sub-variety of An for which π −1 (t0 ) = {t0 } × Vt0 holds.
The polynomials q, w1 , . . . , wn define a geometric solution of the zero
dimensional variety Vt0 with respect to the linear form U .
47
Assume we are given the straight–line program β computing F1 , . . . , Fn , the
coordinates defining U , the point t0 and the vectors that stand for the dense
representation of q, v1 , . . . , vn . Then, we can compute a geometric solution
of V with O((nT + n3 )M(deg π)M(degT mu )) operations in K.
Proof. We want to compute polynomials q̂, v̂1 , . . . , v̂n ∈ K(T )[Y ] defining
a geometric solution of V with respect to the linear form U . Let E :=
degT mu . In order to do this, we claim that we can compute polynomials
Q, W1 , . . . , Wn ∈ K[T, Y ] such that
Q(T, U (X)) ≡ q̂(T, U (X))
Wi (T, U (X)) ≡
∂ q̂
∂Y
(T, U (X))
mod (T − t0 )E
−1
and
(2.3)
· v̂i (T, U (X))
mod (T − t0 )
E
in K[[T − t0 ]][X]/(F1 , . . . , Fn ) for 1 ≤ i ≤ n. Indeed, the existence of
q̂, v̂1 , . . . , v̂n follows from our hypotheses. Since U separates the points of
π −1 (t0 ) we conclude that q̂(t0 , Y ) is a separable element of K[Y ]. This shows
that q̂(t0 , Y ) and ∂ q̂(t0 , Y )/∂Y are relatively prime in K[Y ]. Hence, ∂ q̂/∂Y
is invertible mod q̂(T, Y ) in K[[T − t0 ]][Y ] and thus, ∂ q̂/∂Y (T, U (X)) is invertible in K[[T − t0 ]][X1 , . . . , Xn ]/(F1 , . . . , Fn ).
As shown in the next claim, we compute the polynomials W1 , . . . , Wn , Q ∈
K[T, Y ] through a recursive procedure which at step κ ∈ {0, 1, . . . , ⌈log E +
[κ]
[κ]
1⌉} produces polynomials R1 , . . . , Rn+1 that agree with (∂ q̂/∂Y )−1 v̂1 , . . . ,
κ
(∂ q̂/∂Y )−1 v̂n , q̂ in K[[T − t0 ]][Y ] mod (T − t0 )2 , e.g., with precision 2κ , as
depicted in (2.3). Once we computed the approximations with precision 2E,
we can use Padé approximation to recover the sought geometric solution.
Let A := K[[T − t0 ]] and o := (T − t0 ). Write Xn+1 := Y and for the
remainder of this section let X := (X1 , . . . , Xn+1 ). Let
ϕ(X) := (X1 − W1 (Xn+1 ), . . . , Xn − Wn (Xn+1 ), Q(Xn+1 ))
in (A[X])n+1 denote the vector whose coordinates are the images of
X1 − (∂ q̂/∂Y )−1 v̂1 , . . . , Xn − (∂ q̂/∂Y )−1 v̂n , q̂
in A[X]. Write ϕ(X) = (ϕ1 , . . . , ϕn+1 )(X). Note that, for every 1 ≤ j ≤
n + 1, the polynomial ϕj of A[X]:
48
is monic in Xj ,
is reduced modulo (ϕj+1 , . . . , ϕn+1 ) in the lexicographical order with
Xn+1 > Xn > · · · > X1 .
depends only on the variables Xj , . . . , Xn+1 .
Moreover, let
F := (F1 (X), . . . , Fn (X), Xn+1 − U (X))
denote the above tuple of polynomials considered as elements of A[X].
Claim: We claim that for each κ ∈ Z≥0 , we can compute polynomials
κ+1
[κ]
[κ]
R1 , . . . Rn+1 in K[T, X] such that their images in (A/o2 )[X] satisfy the
following conditions:
[κ]
a) Rj is monic in Xj ,
[κ]
κ
b) Rj ≡ ϕj mod o2 in (A/o2
κ+1
)[X], and
[κ]
c) Rj has the same support as ϕj
for 1 ≤ j ≤ n + 1. (The tuple R[κ] is to be understood as an approximation
κ+1
to ϕ of precision 2κ in the modular ring A/o2 [X].)
Proof of the Claim. Consider the ideals
(F1 (t0 , X), . . . , Fn (t0 , X), Xn+1 − U (X)),
(X1 − W1 (t0 , Xn+1 ), . . . , Xn − Wn (t0 , Xn+1 ), Q(t0 , Xn+1 ))
(2.4)
(2.5)
of K[X]. Since the ideal spanned by the (F1 (t0 , X), . . . , Fn (t0 , X)) in K[X]
is radical, the ideal in (2.4) is also radical. By definition of geometric solution, the varieties generated by the ideals (2.4) and (2.5) agree. Hence, it
follows that these ideals are equal. Considering the univariate polynomials
q, w1 , . . . , wn ∈ K[Xn+1 ] in the given geometric solution of π −1 (t0 ) as polynomials in A[X], it follows that Q(t0 , Xn+1 ) = q(Xn+1 ) and Wi (t0 , Xn+1 ) =
wi (Xn+1 ) for 1 ≤ i ≤ n in A[X]. Then, each entry of
R[0] := (X1 − w1 (Xn+1 ), . . . , Xn − wn (Xn+1 ), q(Xn+1 ))
49
0
is congruent to the corresponding entry of ϕ modulo (T − t0 ) = o2 in A[X]
which implies b) for κ = 0. Conditions a) and c) are also fulfilled for κ = 0
by definition.
Let κ ≥ 0 and assume that R[κ] fulfilling the hypotheses has been computed.
κ+1
Let Hκ denote the ring Hκ := (A/o2 )[X]/(R[κ] ). By construction, Hκ
κ+1
αn+1
is a free (A/o2 )–module, and the monomials {X1α1 · · · Xn+1
: 0 ≤ αi <
[κ]
degXj Rj for 1 ≤ j ≤ n + 1} form a basis of this module. For any element
a ∈ A[X], write aκ for its projection in Hκ .
Let J(F ) := (∂Fi /∂Xj )1≤i,j≤n+1 denote the Jacobian matrix of F1 , . . . , Fn+1
with respect to X1 , . . . , Xn+1 in A[X]. We make the following sub-claim:
Sub-claim 1. It holds that
J(R[κ] )κ J(F )−1
κ F κ = ϕκ
in Hκ .
Proof of sub-claim 1. Note that the Jacobian determinant det(J(F )) is invertible in A[X]/(ϕ), since it is invertible in K(T )[X]/(ϕ) and π is unramified
at t0 .
Note that the ideals
(F1 (T, X), . . . , Fn (T, X), Xn+1 − U (X))
and
(X1 − W1 (T, Xn+1 ), . . . , Xn − Wn (T, Xn+1 ), Q(T, Xn+1 ))
of K(T )[X] are equal. This implies that we may write every coordinate
in F as a combination of the coordinates of ϕ. Moreover, we can assume
that none of the denominators in T appearing in these equalities vanishes at
T = t0 . This implies that there exists a matrix A ∈ A[X](n+1)×(n+1) such
that F = Aϕ.
Since F = Aϕ in A[X] holds, the equality
J(F ) = AJ(ϕ) + B
50
follows, where B ∈ A[X](n+1)×(n+1) is a matrix whose entries belong to the
ideal (ϕ1 , . . . , ϕn+1 ) ⊂ A[X]. On the other hand, by hypotheses we have that
κ
κ+1
[κ]
ϕj ≡ Rj mod o2 holds in (A/o2 )[X] and hence
ϕj ≡ 0 mod o2
κ
κ+1
in (A/o2 )[X]/(R[κ] ) = Hκ . Combining these two facts, we deduce that in
the equality
J(F )κ = Aκ J(ϕ)κ + Bκ
(2.6)
κ
of Hκ , all the entries in Bκ belong to the ideal o2 Hκ , or equivalently, that
mod o2
J(F )κ = Aκ J(ϕ)κ
κ
in Hκ .
κ
Since J(F )κ is invertible in Hκ /o2 Hκ , this implies that Aκ and J(ϕ)κ
κ
are also invertible in Hκ /o2 Hκ . Moreover, by Hensel’s Lemma we deduce
that J(F )κ , Aκ and J(ϕ)κ are also invertible in Hκ .
By multiplying both sides of (2.6) to the right by J(ϕ)−1
κ and to the left
−1
by J(ϕ)κ J(F )κ in Hκ we have
−1
−1
I = J(ϕ)κ J(F )−1
κ Aκ + J(ϕ)κ J(F )κ Bκ J(ϕ)κ ,
which in turn, implies
J(ϕ)κ J(F )−1
κ Aκ = I + C κ
(2.7)
κ
where the entries of Cκ belong to the ideal o2 of Hκ . Finally, multiplying
both sides of (2.7) to the right by ϕκ and replacing Aκ ϕκ by Fκ (they are
equal!), we obtain
−1
J(ϕ)κ J(F )−1
κ Aκ ϕκ = J(ϕ)κ J(F )κ Fκ = ϕκ + Cκ ϕκ = ϕκ
κ
in Hκ , where Cκ ϕκ = 0 since they are both elements of o2 and therefore
κ+1
their product belongs to o2 .
κ
In fact, since R[κ] −ϕκ is in the ideal o2 Hκ of Hκ , it follows that each entry
κ
of J(R[κ] ) − J(ϕκ ) is also in o2 Hκ . On the other hand, every entry of Fκ =
κ
Aκ ϕκ is also in this ideal, since every entry in ϕκ belongs to o2 Hκ . Hence,
combining these two facts we deduce that (J(R[κ] ) − J(ϕκ ))J(F )−1
κ Fκ = 0 in
51
Hκ and therefore we may replace J(ϕκ ) by J(R[κ] ) in the above equality and
obtain:
J(R[κ] )κ J(F )−1
(2.8)
κ F κ = ϕκ
in Hκ which proves sub-claim 1. QED.
Assume J(R[κ] )κ J(F )−1
κ Fκ is computed in Hκ and let δκ = (δκ,1 , . . . , δκ,n )
2κ+1
denote a representative of J(R[κ] )κ J(F )−1
[X] that it is chosen
κ Fκ in A/o
so that1
[κ]
for 1 ≤ i, j ≤ n + 1.
(2.9)
degXi (δκ,j ) < degXi Ri
From (2.8) we deduce that each entry in the tuple δκ −ϕκ is in the ideal (R[κ] )
κ+1
of A/o2 [X]. Evidently the same happens with the entries of δκ − ϕκ + R[κ] ;
κ+1
[κ]
moreover, the next sub-claim proves that Rj + δκ,j = ϕj,κ in A/o2 [X].
[κ]
Sub-claim 2. It holds that δκ,j = ϕj,κ − Rj in A/o2
[κ]
κ+1
[X].
Proof of sub-claim 2. Let gj := δκ,j − ϕj,κ + Rj in A/o2
1 ≤ j ≤ n + 1.
κ+1
[X] for every
[κ]
When j = n+1 we have (by the inductive hypotheses) that both Rn+1 and
ϕn+1,κ depend only on Xn+1 , they are monic and of the same degree, so that
[κ]
[κ]
degXn+1 (Rn+1 −ϕn+1 ) < degXn+1 (Rn+1 ). Moreover, from (2.9) we deduce that
[κ]
[κ]
degXi (gn+1 ) < degXi (Ri ) for every 1 ≤ i ≤ n + 1. Since degXi (Ri ) = 1 for
κ+1
1 ≤ i ≤ n, we see that gn+1 ∈ A/o2 [Xn+1 ]. Furthermore, degXn+1 gn+1 <
κ+1
[κ]
[κ]
degXn+1 Rn+1 and gn+1 is in the ideal (Rn+1 ) of A/o2 [Xn+1 ]. This proves
that gn+1 = 0.
[κ]
For j ≤ n, as before, we have that both Rj and ϕj,κ are monic in Xj
[κ]
[κ]
[κ]
and of the same degree degXj (Rj ), so that degXj (Rj − ϕj,κ ) < degXj (Rj ).
[κ]
Moreover, from (2.9) we deduce that degXj (gj ) < degXj (Rj ).
Since ϕj is reduced modulo ϕj+1 , . . . , ϕn+1 , this implies that
[κ]
for j < i ≤ n + 1.
degXi (ϕj ) < degXi (Ri )
[κ]
Hence, this also holds for ϕj,κ and for Rj ; in the latter case this is because,
[κ]
by hypotheses, Rj has the same support as ϕj . Again, by equality (2.9) we
1
We take the convention that the degree of the polynomial zero deg 0 = −∞ is smaller
than every integer; so that when δκ,j = 0 the inequality holds.
52
[κ]
deduce that degXi (gj ) < degXi (Ri ) for i = j + 1, . . . , n + 1. Finally, since
[κ]
Rj and ϕj,κ have the same support and only depend on Xj , . . . , Xn+1 we
κ+1
deduce as before that gj = 0 in A/o2 [X] thus proving sub-claim 2. QED.
κ+2
Thence, we can set R[κ+1] as a representative of R[κ] + δκ in A/o2 [X].
One verifies that R[κ+1] fulfils all the conditions in the inductive argument
proving our claim. QED.
To estimate the complexity of each step, we first make some auxiliary
estimations. The cost of operations in Hκ requires computations modulo
κ+1
κ+1
R[κ] and next reduction modulo o2
= (T − t0 )2 . Since the polynomials
[κ]
[κ]
R1 , . . . , Rn have their total degree bounded by deg π in X1 , . . . , Xn+1 , any
arithmetic operation in Hκ has a cost of roughly O(n(deg π)2κ+1 ) arithmetic
operations in K. Since R[κ] is constituted of n + 1 vectors in Hκ , evaluating
the vector has a cost of O(n(deg π)2κ ) arithmetic operations in K. The evaluation of the matrix J(F ) is done using Baur-Strassen’s algorithm ([BS83])
and then the evaluation of F and J(F ) can be estimated by O(nT). Finally,
linear algebra can be done using Samuelson’s algorithm in O(n4 ) complexity.
Combining these facts we deduce that roughly O((nT+n4 )(deg π)2κ+1 ) arithmetic operations in K are required for each recursive step. Therefore, the
computation of the approximation up to precision 2E = 2 degT mu requires
roughly O((nT + n4 )(deg π) degT mu ) arithmetic operations in K.
The reconstruction of the sought geometric solution is done using Padé
approximation and uses at most O(n degT mu ) additional operations in K
(cf. [BP94], [vzGG99]), thus the estimate for the whole procedure remains
in O((nT + n4 )(deg π) degT mu ) arithmetic operations in K.
Let t1 ∈ K be a new point in A1 . Then, we can compute the geometric
solution of π −1 (t1 ) by first applying the above procedure, specialising in t1
and eventually cleaning multiplicities.
Lemma 2.3 (see [GLS01, §6.5]). Let notations and assumptions be as in
Theorem 2.2. Let be given t1 ∈ K such that U separates the points of the
fibre π −1 (t1 ) and a geometric solution of V with underlying linear form U .
Then we can compute a geometric solution of π −1 (t1 ) with O(nM(deg π))
arithmetic operations in K.
2.4.2.
53
From minimal equations to geometric solutions
We have exhibited a lifting technique which, given polynomials defining a
curve, a lifting point, the coefficients of a linear separating form and the geometric solution of the corresponding fibre, computes the geometric solution
of such a curve. However, in some cases it is possible to compute the minimal
equation of the image of a projection through a different procedure and what
is needed is to complete this to a geometric solution (e.g., see Section 5.2.4).
We describe an algorithm that, given a procedure for computing a minimal
equation for a primitive element in a zero-dimensional or a one-dimensional
equidimensional variety, produces the geometric solution of this variety.
This procedure is extracted from [JMSW09] and [DMW09], both of which
I co-authored, and is inspired in an idea of [GLS01].
Let notions and notations be as above and let Λ1 , . . . , Λn be new indeterminates over K. We return to the notation X := (X1 , . . . , Xn ) and
set Λ := (Λ1 , . . . , Λn ). Denote by IK(Λ) ⊂ K(Λ)[X] the extension ideal of
I(V ) ⊂ K[X]. Note that the linear form U := Λ1 X1 + . . . + Λn Xn separates the points of V (IK(Λ) ). Suppose that we are given an algorithm Ψ
in K(Λ) which computes the minimal polynomial mU ∈ K[Λ, Y ] of U in
K(Λ) → K(Λ)/IK(Λ) with T arithmetic operations in K(Λ) and a separating
linear form u := λ1 X1 + . . . + λn Xn ∈ K[X] such that the vector (λ1 , . . . , λn )
annihilates none of the denominators in K[Λ] of any intermediate result of
the algorithm Ψ.
Note that mU (Λ, U ) is in I(An × V ). Since I(An × V ) is generated by
polynomials in K[X], it follows that the derivative of mU (Λ, U ) with respect
to Λi ,
∂mU
∂mU
∂mU
(Λ, U )Xi +
(Λ, U ) =
(Λ, U ),
Λi
∂Y
∂Λi
also belongs to I(An × V ) for 1 ≤ i ≤ n. Observe that the equality deg mu =
D := #V and the estimate deg((∂mU /∂Λi )(λ, Y )) ≤ D−1 hold. Substituting
λ for Λ in mU (Λ, Y ) we obtain mu (Y ). Making the same substitution in
(∂mU /∂Y )(Λ, Y )Xi + (∂mU /∂Λi )(Λ, Y ) for 1 ≤ i ≤ n and reducing modulo
mu (Y ) we obtain polynomials
∂mu
(Y )Xi − vi (Y )
∂Y
that belong to the ideal I(V ) of K[X].
54
Since mu (Y ) is square-free, it follows that mu (Y ) and (∂mu /∂Y )(Y ) are
relatively prime in K[Y ]. Thus, we can compute a polynomial (∂mu /∂Y )−1 (Y )
of degree at most D − 1 representing the inverse of (∂mu /∂Y )(Y ) modulo mu (Y ). Hence, multiplying this polynomial by vk modulo mu (Y ) we
obtain a polynomial wk (Y ) := (∂mu /∂Y )−1 vk such that deg wk < D and
wk (u) − Xk ∈ I(V ) for 1 ≤ k ≤ n. Since u is a separating linear form, it
follows that w1 , . . . , wn together with mu form a geometric solution of V .
In order to compute the polynomials w1 , . . . , wn , we observe that the
Taylor expansion of mU (Λ, Y ) in powers of Λ−λ has the following expression
mU (Λ, Y ) = mu (Y ) +
n X
∂mu
k=1
∂Y
(Y )Xk − vk (Y ) (Λk − λk ) mod(Λ − λ)2 .
(2.10)
We compute this first order expansion by computing the first-order Taylor
expansion of each intermediate result in the algorithm Ψ. In this way, each
arithmetic operation in K(Λ) arising in the algorithm Ψ becomes an arithmetic operation between two polynomials of K[Λ] of degree at most 1, and
is truncated up to order (Λ − λ)2 . Since the first-order Taylor expansion of
an addition, multiplication or division of two polynomials of K[Λ] of degree
at most 1 requires O(n) arithmetic operations in K, we have that the whole
step requires O(nT) arithmetic operations in K, where T is the number of
arithmetic operations in K(Λ) performed by the algorithm Ψ.
The computation of the polynomials w1 , . . . , wn requires the inversion of
∂mu /∂Y modulo mu (Y ) and the modular multiplication
wk (Y ) := (∂mu /∂Y )−1 vk (Y )
for 1 ≤ k ≤ n. These steps can be executed with additional O nM(D)
arithmetic operations in K.
Hence, this solution can be computed with O(n(T + M(D))) arithmetic
operations in K. We can thus state this in the form of the following lemma.
Lemma 2.4. Suppose that we are given:
1. an algorithm Ψ in K(Λ) which computes the minimal polynomial mU ∈
K[Λ, Y ] of U := ΛX1 + · · · + Λn Xn with T arithmetic operations in
K(Λ),
55
2. a separating linear form u := λ1 X1 + · · · + λn Xn ∈ K[X] such that the
vector (λ1 , . . . , λn ) annihilates none of the denominators that are part
of the intermediate results of the algorithm Ψ in K[Λ].
Then a geometric solution of the variety V can be (deterministically) computed with O n(T + M(D)) arithmetic operations in K.
In the situations of interest T is a polynomial of degree at least quadratic
in D. Then the term O(nT) dominates the complexity of the whole procedure
for computing the geometric solutions under consideration, which includes
the cost of the computation of projections and the subsequent use of this
procedure in order to obtain such a geometric solution. Hence, we shall
concentrate in computing projections efficiently.
56
Chapter 3
Robust algorithms for generalized
Pham systems
We announced earlier that this chapter was going to be devoted to a family called generalized Pham systems or strict complete intersections, which
were introduced in [PS04] and [CDS96]. Recall that an n–dimensional generalized Pham system is defined by n polynomials of the form φi − ϕi (1 ≤
i ≤ n), where φi ∈ Q[X1 , . . . , Xn ] is homogeneous of degree di , ϕi has degree
less than di and φ1 , . . . , φn define the empty projective variety of Pn−1 (C).
This chapter is based on an article of the same title that I co-authored
with Ezequiel Dratman and Guillermo Matera ([DMW09]) and has two main
results, one on lower bounds for generalised Pham systems and one on upperbounds. We shall cover only the upper bound and not include the lower
bound which is certainly not the topic of this thesis.
In Section 3.2 we prove Theorem 3.15. Namely, we exhibit a probabilistic
algorithm which solves generalized Pham systems with quadratic complexity
in the Bézout number D = d1 · · · dn . As we discussed, the straightforward
application of the deformation technique of Theorem 2.2 does not yield an
efficient algorithm. A clever procedure that uses multiple deformations will
endow us with a better result.
For this purpose, for 1 ≤ r ≤ n + 1, we define a sequence of homotopies
πr : Vr → A1 such that:
1. The fibre π1−1 (0) is unramified and “easy to solve”,
57
58
CHAPTER 3. GENERALIZED PHAM SYSTEMS
2. For 1 ≤ r ≤ n, the fibre πr−1 (1) is unramified and the equality πr−1 (1) =
−1
−1
(1) = {1} × V holds, where V = V (φ1 −
πr+1
(0) holds, and πn+1
ϕ1 , . . . , φn − ϕn ).
3. For 1 ≤ r ≤ n + 1, our projection algorithm lifts πr−1 (0) to Vr with
complexity that is quadratic in the Bézout number of the original input
system.
These homotopies are reminiscent of certain “piecewise–linear homotopies” of
numerical continuation methods acting coordinate by coordinate (see, e.g.,
[Sai83], [Duv90]). Basing P
our algorithm in
these deformations, we shall take
roughly O (nT+n3 )M(D) n+1
M(D/d
)
arithmetic operations in Q to comr
r=1
pute the geometric solution of a generalized Pham system where T is the
number of arithmetic operations required by the straight-line program that
computes φ1 − ϕ1 , . . . , φn − ϕn .
Comparison with related work
Let f1 , . . . , fn ∈ Q[X1 , . . . , Xn ] be the input polynomials. Since the input
system f1 = · · · = fn = 0 has no points at infinity, deterministic algorithms
for solving zero–dimensional homogeneous systems can be applied to the
homogenizations of f1 , . . . , fn . Such ideas were applied in [Laz83] to deterministically solve zero–dimensional systems f1 = · · · = fn = 0 having at
most a finite number of points at infinity with complexity of order DO(1)
(see also [Giu89], [Giu91] for similar results and [Bar04] for the complexity
of deterministic algorithms for systems defined by a regular sequence).
In connection with the complexity of probabilistic algorithms solving generalized Pham systems, we observe that our algorithm solves any generalized
Pham system with quadratic complexity in the Bézout number D = d1 · · · dn ,
extending thus the results of [MP00], [MT00] and [BMWW04, Section 5] to
generalized Pham systems.
Another probabilistic algorithm solving generalized Pham systems with
quadratic complexity in the Bézout number D is obtained by a clever application of the algorithm of [GLS01]. Indeed, since a generalized Pham system
is not necessarily defined by a reduced regular sequence, this inhibits the
straightsforward application of the algorithm of [GLS01]. Nevertheless, if
g1 , . . . , gn ∈ Q[X1 , . . . , Xn ] are n generic linear combinations of f1 , . . . , fn ,
59
then with high probability the polynomials g1 , . . . , gn−1 form a reduced regular sequence and g1 , . . . , gn define the same variety as f1 , . . . , fn (see, e.g.,
[KP96, Section 6]). In such a case, the application of the algorithm of [GLS01]
to the polynomials g1 , . . . , gn has quadratic complexity in the Bézout number
D = d1 · · · dn , rather than in max{d1 , . . . , dn }n , as a simple minded analysis
might suggest.
We observe that any generalized Pham system can be (partially) solved
applying the non–universal symbolic homotopy algorithm of [PS04]. This
algorithm has a cost which is roughly of order D∗2.38 , where D∗ denotes the
degree of certain irreducible component of the curve introduced by the linear homotopy considered by the authors. As a consequence, the resulting
algorithm will improve over our estimates for certain particular systems for
which the irreducible component of the underlying curve under consideration
has “low” degree. Nevertheless, for a generic generalized Pham system such a
curve is irreducible of degree D, which implies that the cost of the algorithm
of [PS04] is roughly of order O(D2.38 ).
When considering the particular families of generalized Pham systems
defined below in Section 3.1.2 one is often interested only in the real positive
solutions ([Can84, §20.3], [Pao92a, §1.1]). It has been shown that the set of
stationary solutions of the heat equation with monomial reaction terms and
boundary conditions has only one positive solution in the cases of “small” or
“large” absorption ([CFQ91a]). Furthermore, in [DM09], [Dra10] it has been
shown that there are no spurious positive stationary solutions of the corresponding semidiscretization, namely positive solutions not corresponding to
anyone of the heat equation. This implies that there is only one positive solution of the systems arising from the semidiscretization in the cases mentioned
above, provided that the number of nodes involved in the semidiscretization
is large enough. Hence, the specially-crafted algorithms of [DM09, Dra10]
compute ε–approximations of these solutions in time that is linear in n, and
thus are more efficient than ours, which compute all the complex solutions
and perform O(d2n ) arithmetic operations. In connection with this matter,
it may also be worthwhile to mention that a generalization of this result has
been obtained in [Dra13a], [Dra13b].
60
3.1.
A catalogue of generalized Pham systems
In this section we discuss some sources of interest for the notion of a generalized Pham system. For this purpose we introduce three particular classes
of zero–dimensional generalized Pham systems: Pham systems, systems arising in the analysis of the stationary solutions of certain parabolic differential
equations and generalized Reimer systems.
3.1.1.
Pham Systems
Fix n, d1 , . . . , dn ∈ N. Let g1 , . . . , gn ∈ Q[X1 , . . . , Xn ] be polynomials
with deg(gi ) < di for 1 ≤ i ≤ n, and consider the polynomials
f1 := X1d1 − g1 , . . . , fn := Xndn − gn .
The map f := (f1 , . . . , fn ) : Cn → Cn is called a Pham map in [VA85, Chapter 1, Section 5.2] and is considered in connection with the study of the local
multiplicity of a holomorphic map. Consistently, the system f1 = 0, . . . , fn =
0 is called a Pham system (see, e.g., [GLGV98], [MP00], [BMWW04]).
3.1.2.
Systems Coming from a Semidiscretization of certain Parabolic Differential Equations
Let Z be an indeterminate and let f, g, h ∈ Q[Z] be given polynomials. Several problems concerning unidimensional nonlinear heat transfer are
described by a partial differential equation of the form ut = f (u)xx + g(u)
in a bounded domain, say (0,
1) × [0, t0 ), with (Newmann) boundary conditions f (u)x (1, t) = h u(1, t) and f (u)x (0, t) = 0 in [0, t0 ) and u(x, 0) ≥ 0
in [0, 1] (see, e.g., [Pao92b]). In particular, the asymptotic behavior of the
solutions of such boundary value problems has been intensively analyzed (cf.
[SGKM95]). This behavior is mainly described by the corresponding stationary solutions, namely, the positive solutions of the ordinary differential
equation 0 = f (u)′′ + g(u) with boundary conditions f (u)′ (1) = h u(1) and
f (u)′ (0) = 0.
The usual numeric treatment of this latter problem consists in finding a
numerical approximation provided by a standard second order finite difference scheme (see, e.g., [BR01], [FGR02]). The solutions of such numerical
3.1. A CATALOGUE OF GENERALIZED PHAM SYSTEMS
61
approximation are represented by the system defined by the following polynomials:
f1 := 2(n − 1)2 f (X2 ) − f (X1 ) − g(X1 ), fi := (n − 1)2 f (Xi+1 ) − 2f (Xi ) +
f (Xi−1 ) − g(Xi ), (2 ≤ i ≤ n − 1)
fn := 2(n − 1)2 f (Xn−1 ) − f (Xn ) + 2(n − 1)h(Xn ) − g(Xn ).
(3.1)
Two important cases of study are the porous medium equation with nonlinear source terms and boundary condition (see, e.g., [Hen81], [Pao92b]), which
leads to instances of (3.1) with f = h := Z d and g := Z (see, e.g., [FGR02]),
and the heat equation with polynomial reaction terms and boundary conditions, which leads to instances of (3.1) with f := Z, h := Z d1 and g := Z d2
(see, e.g., [BR01], [MD04], [MD05]).
3.1.3.
Reimer Systems
We now consider another family of examples called (generalized) Reimer
systems (see [BM96], [BMWW04]). A generalized Reimer system is defined
by polynomials f1 , . . . , fn ∈ Q[X1 , . . . , Xn ] of the following form:
fi := αi +
n
X
ai,j Xji+1 ,
(3.2)
j=1
where ai,j , αi (1 ≤ i, j ≤ n) are suitable elements of Q with αi 6= 0 for
1 ≤ i ≤ n. More precisely, in [BMWW04, Lemma 17] it is shown that there
2
exists a nonempty Zariski open set U ⊂ Cn with the following property: for
every a := (ai,j )1≤i,j≤n ∈ U, the corresponding polynomials f1 , . . . , fn in (3.2)
define a zero–dimensional system with (n + 1)! distinct complex solutions. A
system f1 = · · · = fn = 0 is called a generalized Reimer system if f1 , . . . , fn
are defined as in (3.2) with αi 6= 0 for 1 ≤ i ≤ n and a := (ai,j )1≤i,j≤n ∈ U.
3.1.4.
Generalized Pham Systems
Let f1 , . . . , fn ∈ Q[X1 , . . . , Xn ] be given polynomials of (total) positive
degrees d1 , . . . , dn respectively. Following [PS04] we say that f1 , . . . , fn define a generalized Pham system if the projective variety {x̄ ∈ Pn−1 (C) :
62
φ1 (x̄) = 0, . . . , φn (x̄) = 0} is empty, where φi ∈ Q[X1 , . . . , Xn ] denotes the
homogeneous component of fi of degree di for 1 ≤ i ≤ n.
It is easy to see that the systems introduced in Sections 3.1.1, 3.1.2 and
3.1.3 are generalized Pham systems. We remark that the solution set of a
generalized Pham system is a zero–dimensional affine variety of Cn (see, e.g.,
[PS04, Proposition 18]).
3.2.
The Solution of a Generalized Pham System
Let f1 , . . . , fn be polynomials in Q[X] which can be computed by a
division–free straight–line program of length T, and let V := V (f1 , . . . , fn )
denote the affine subvariety of An defined by f1 , . . . , fn . For 1 ≤ i ≤ n, set
di := deg fi and write fi = φi + ϕi , where φi ∈ Q[X] is the (nonzero) homogeneous component of fi of degree di . Let δ := deg V denote the degree of
V . Finally, set D := d1 · · · dn and note that δ ≤ D by the Bézout inequality.
Assume that f1 , . . . , fn define a generalized Pham system, that is, the
projective variety {φ1 (x̄) = 0, , . . . , φn (x̄) = 0} ⊂ Pn−1 is the empty set. The
following result will be important in the sequel.
Lemma 3.1 ([PS04, Proposition 18]). The set of solutions of an n–variate
generalized Pham system is a zero–dimensional affine subvariety of An .
Let f1h , . . . , fnh denote the homogenizations of f1 , . . . , fn with homogenizing variable X0 . Observe that the polynomials f1h , . . . , fnh are of the form
f1h = φ1 (X) + X0 ϕ
e1 (X0 , X), . . . , fnh = φn (X) + X0 ϕ
en (X0 , X),
for some polynomials ϕ
e1 , . . . , ϕ
en ∈ Q[X0 , X]. From this representation we
deduce that the projective variety V h := {f1h (x̄) = 0, . . . , fnh (x̄) = 0} ⊂ Pn
is contained in the Zariski–open set {x0 6= 0}, or equivalently the ideal generated by X0 is not contained in the ideal generated by f1h , . . . , fnh . By the
Bézout theorem in the form of [EH99, Theorem III.71] it follows that the projective variety V h has precisely D points in Pn , counting multiplicities. Since
V h has no points at infinity, from [CGH91, Proposition 1.11] we conclude
that V has D points, counting multiplicities.
3.2. THE SOLUTION OF A GENERALIZED PHAM SYSTEM
3.2.1.
63
The architecture of our solution method
We introduce a sequence of deformations which play the role of certain
piecewise–linear homotopies of numerical continuation methods acting coordinate by coordinate (cf. [Sai83], [Duv90]). More precisely, for 1 ≤ r ≤ n + 1
we introduce a deformation πr : Vr → A1 such that the following conditions
are satisfied:
(i) Vr ⊂ An+1 is equidimensional of dimension 1 for 1 ≤ r ≤ n + 1,
(ii) πr is dominant and generically unramified for 1 ≤ r ≤ n + 1,
(iii) deg πr = D = #πr−1 (0) and deg Vr ≤ 2D holds for 1 ≤ r ≤ n + 1,
−1
(iv) πr−1 (1) = πr+1
(0) for 1 ≤ r < n + 1,
−1
(v) π1−1 (0) is “easy to solve” and πn+1
(1) = {1} × V holds.
−1
In order to compute a geometric solution of the fibre πn+1
(1) = {1} × V ,
and thus of V , we shall apply repeatedly the “projection algorithm” of Theorem 2.2. This algorithm takes as input a geometric solution of the unramified
−1
fibre πr+1
(0) = πr−1 (1) of the morphism πr+1 and outputs a geometric solution
of Vr+1 for any 1 ≤ r ≤ n. Making the substitution T = 1 in the polynomials
that form the computed geometric solution of Vr+1 we obtain polynomials
−1
−1
that form a geometric solution of πr+1
(1) = πr+2
(0). Since the fibre π1−1 (0)
is easy to solve, after n applications of such a projection algorithm we obtain
−1
a geometric solution πn+1
(1).
Assume that we are given deformations πr : Vr → A1 (1 ≤ r ≤ n + 1)
satisfying conditions (i)–(v) above. The following is a sketch of our algorithm
for computing a geometric solution of the input system f1 = 0, . . . , fn = 0.
Algorithm 3.2 (Sketch of the algorithm for solving f1 = 0, . . . , fn = 0).
1. Find a geometric solution of the “easy–to–solve” fibre π1−1 (0).
2. For r = 1 to n do:
a) Apply a “projection algorithm” in order to compute a geometric
solution of Vr from the geometric solution of πr−1 (0) computed in
the previous step.
64
b) Make the substitution T = 1 in the polynomials that form the
geometric solution of Vr computed in the previous step. These
−1
polynomials form a geometric solution of πr+1
(0) = πr−1 (1) for
1 ≤ r ≤ n − 1 and may include multiplicities for r = n.
3. Clean multiplicities from the polynomials computed in the previous step
for r = n to obtain a geometric solution of Vn+1 = {1} × V , and thus
of V .
3.2.2.
Designing suitable deformations
In the description of our deformations πr : Vr → A1 we shall make use of
certain Q–linearly independent linear forms Y1 , . . . , Yn ∈ Q[X] and nonzero
rational numbers b1 , . . . , bn ∈ Q \ {0} to be fixed during the setup stage of
our algorithm (Section 3.2.3 below).
Fix r with 1 ≤ r ≤ n and consider the following polynomials of Q[T, X]:

(r)

 Fj (T, X) := φj (X) + bj (1 ≤ j ≤ r − 1),
(r)
(3.3)
Fr (T, X) := Yrdr + T (φr (X) − Yrdr ) + br ,

 F (r) (T, X) := Y dj + b (r + 1 ≤ j ≤ n).
j
j
j
In particular, for r = 1 we have
(1)
F1 (T, X) := Y1d1 + T (φ1 (X) − Y1d1 ) + b1 ,
d
Fj (1) (T, X) := Yj j + bj
for 2 ≤ j ≤ n
and for r = n we have
Fj (n) (T, X) := φj (X) + bj
for 1 ≤ j ≤ n − 1,
(n)
dn
dn
Fn (T, X) := Yn + T (φn (X) − Yn ) + bn .
Finally, consider the following polynomials of Q[T, X]:
(n+1)
Fi
(T, X) := φi (X) + T (ϕi (X) − bi ) + bi
(r)
for 1 ≤ i ≤ n.
(r)
(3.4)
For any 1 ≤ r ≤ n + 1, let Ir := (F1 , . . . , Fn ) ⊂ Q[T, X], let Jr :=
(r)
(r)
(r)
det(∂Fi /∂Xj )1≤i,j≤n be Jacobian determinant of F1 , . . . , Fn with respect
to the variables X and let Vr := V (Ir : Jr∞ ) ⊂ An+1 be the variety defined by
65
the saturation (Ir : Jr∞ ). The rth deformation πr : Vr → A1 is determined
by the projection πr (t, x) := t and the affine variety Vr ⊂ An+1 .
Having defined the deformations πr : Vr → A1 (1 ≤ r ≤ n + 1), we discuss
the validity of conditions (i)–(v) of the previous section. The fulfillment of
these conditions relies on a suitable choice of the coefficients of the linear
forms Y1 , . . . , Yn and the rational numbers b1 , . . . , bn . In the next section
we prove that for certain random choice of these coefficients, the following
assertions hold with high probability:
(r)
(A) The polynomials Fi (t, X) (1 ≤ i ≤ n) define a generalized Pham
system for a generic choice t ∈ A1 and every 1 ≤ r ≤ n + 1 (see
Corollary 3.8 bellow).
(r)
(r)
(B) The affine variety V F1 (0, X), . . . , Fn (0, X) consists of D nonsingular points for 1 ≤ r ≤ n + 1 (see Proposition 3.9 below).
Assuming that assertions (A)–(B) hold, from (A) and Lemma 3.1 we
(r)
(r)
conclude that {F1 (t, X) = 0, . . . , Fn (t, X) = 0} is a zero–dimensional
subvariety of An for a generic choice t ∈ A1 . This shows that the generic
fibre of the projection mapping πr : V (I (r) ) → A1 is nonempty for 1 ≤ r ≤
n + 1 (and consists of at most D points by the Bézout inequality (2.1)).
Furthermore, (B) asserts that the fibre πr−1 (0) consists of D nonsingular
points for 1 ≤ r ≤ n + 1. Summarizing, under the assumption of assertions
(A)–(B) it follows that the deformations πr : Vr → A1 satisfy the following
conditions for 1 ≤ r ≤ n + 1:
(C) 0 < #πr−1 (t) ≤ D holds for a generic value t ∈ A1 ,
(D) #πr−1 (0) = D and Jr (0, x) 6= 0 holds for every (0, x) ∈ πr−1 (0).
Proposition 3.3. Let be given polynomials F1 , . . . , Fn ∈ Q[T, X], a positive
integer D and t0 ∈ Q. Let I ⊂ Q[T, X] be the ideal generated by F1 , . . . , Fn
and set W := V (I) = V (F1 , . . . , Fn ). Let J := det(∂Fi /∂Xj )1≤i,j≤n be the
Jacobian determinant of F1 , . . . , Fn with respect to the variables X and let
V := V (I : J ∞ ) ⊂ An+1 be the variety defined by the saturation (I : J ∞ ).
Let π : W → A1 denote the projection π(t, x) := t. Assume that
0 < #π −1 (t) ≤ D holds for a generic value t ∈ A1 ,
66
#π −1 (t0 ) = D and J(t0 , x) 6= 0 holds for every (t0 , x) ∈ π −1 (t0 ).
Then the following assertions hold:
V is an equidimensional variety of dimension 1.
V is the union of all the irreducible components of W having a nonempty
intersection with π −1 (t0 ).
V is the union of all the irreducible components of W projecting dominantly on A1 .
π : V → A1 is a dominant map of degree D.
Proof. First we observe that dim(C) ≥ 1 holds for each irreducible component C of W, since W is defined by n polynomials in an (n + 1)–dimensional
space. By definition, V consists of all irreducible components of W on which
the Jacobian determinant J does not vanish identically.
Let C be an irreducible component of W for which π −1 (t0 ) ∩ C 6= ∅
holds. Consider the restriction π|C : C → A1 of π. Then we have that
π|−1
C (t0 ) is a nonempty variety of dimension zero, which implies that the
generic fibre of π|C is either zero-dimensional or empty. Since dim(C) ≥ 1,
the theorem on the dimension of fibres implies that dim(C) = 1 holds and
that π|C : C → A1 is a dominant map with generically finite fibres. Finally,
[Sha94, §II.6, Theorem 4] shows the inclusion C ⊂ V and, in particular, that
V is nonempty.
Conversely, we have that π −1 (t0 ) ∩ C 6= ∅ holds for any irreducible component C of V. Indeed, assume on the contrary the existence of a component
C0 not satisfying this condition. Let t1 ∈ A1 be a point having a finite fibre
−1
π −1 (t1 ) such that π|−1
C0 (t1 ) and π|C (t1 ) have maximal cardinality for every C
with C ∩ π −1 (t0 ) 6= ∅. This implies that #π −1 (t1 ) > #π −1 (t0 ) = D, leading
to a contradiction.
We conclude that V is the equidimensional variety of dimension 1 which
consists of all the irreducible components C of W with π −1 (t0 ) ∩ C 6= ∅.
Furthermore, this shows that the restriction π|V : V → A1 is a dominant
map of degree D.
Finally we show that V consists of all irreducible components of W projecting dominantly on A1 . Let C be one such irreducible component of W.
67
Then by [Sha94, §II.6, Theorem 4] it follows that there exists an unramified
fibre of π|C . On such a fibre the Jacobian J does not vanish, which in turn
proves that J does not vanish identically on C. Therefore, C ⊂ V holds.
On the other hand, if C is an irreducible component of W for which the
projection π|C : C → A1 is not dominant, then C is the set of common zeros
of the polynomials F1 , . . . , Fn , T − tC for some value tC . Since dim(C) ≥ 1,
we have that the Jacobian matrix ∂(F1 , . . . , Fn , T − tC )/∂(X1 , . . . , Xn , T ) is
singular at every point (x, tC ) of C. Hence, its determinant, which equals J,
vanishes over C.
We claim that the validity of (A)–(B) implies that conditions (i)–(v) of
Section 3.2.1 hold. Indeed, combining (A)–(B) with the first conclusion of
Proposition 3.3 immediately implies that Vr is an equidimensional variety of
dimension 1 for 1 ≤ r ≤ n + 1, that is, condition (i) holds.
By (B) and the third and fourth conclusions of Proposition 3.3 we have
that the morphism πr : Vr → A1 is dominant and generically unramified for
1 ≤ r ≤ n + 1, proving thus that condition (ii) holds.
In the proof of (C)–(D) we have already shown that (A)–(B) imply that
the identity deg πr = πr−1 (0) = D holds for 1 ≤ r ≤ n + 1, which is the first
part of condition (iii). Furthermore, by the Bézout inequality (2.1) we have:
deg V (Ir ) ≤ d1 · · · dr−1 (dr + 1)dr+1 · · · dn ≤ 2D
for 1 ≤ r ≤ n and
deg V (In+1 ) ≤ D.
From these estimates and the definition of the varieties Vr we conclude that
the following estimates, and thus the second part of condition (iii), hold:
deg Vr ≤ d1 d2 · · · dr−1 (dr + 1)dr+1 · · · dn ≤ 2D
for 1 ≤ r ≤ n and
deg Vn+1 ≤ D.
From the second conclusion of Proposition 3.3 we deduce the following
identities, which imply (iv):
# V (Ir ) ∩ {T = 0} = # Vr ∩ {T = 0} = D (1 ≤ r ≤ n + 1),
# V (Ir ) ∩ {T = 1} = # Vr ∩ {T = 1} (1 ≤ r ≤ n).
68
Concerning the second assertion of condition (v), we observe that the
identity V (In+1 ) ∩ {T = 1} = Vn+1 ∩ {T = 1} holds, because In+1 = (In+1 :
∞
Jn+1
) holds since the finiteness of the projection πn+1 : V (In+1 ) → A1 implies
that V (In+1 ) contains no vertical component.
Finally, for the first assertion of (v) we observe that V (I1 ) ∩ {T = 0} is
defined by a “diagonal” square system and therefore can be easily solved (see
the algorithm underlying Lemma 3.11 below). This finishes the proof of our
claim.
3.2.3.
Preparation: random choices
This section is devoted to showing that we can choose the linear forms
Y1 , . . . , Yn and the vector b := (b1 , . . . , bn ) ∈ Qn such that conditions (A)–(B)
above are satisfied.
We prove the technical results that we shall use for establishing conditions
on the coefficients of Y1 , . . . , Yn which assure that the polynomials in (3.3)
and (3.4) define generalized Pham systems. In the sequel we shall use the following simple criterion, which is a well–known consequence of the Macaulay
un-mixedness theorem (see, e.g., [Mat80, Exercise 16.3]).
Remark 3.4. Let Q1 , . . . , Qn ∈ Q[X] := Q[X1 , . . . , Xn ] be nonzero homogeneous polynomials defining the empty projective variety {Q1 (x̄) =
0, . . . , Qn (x̄) = 0} of Pn−1 . Then Q1 , . . . , Qn form a regular sequence in
Q[X].
As a first consequence of this remark we observe that, since f1 , . . . , fn
determine a generalized Pham system, the polynomials φ1 , . . . , φn define the
empty projective variety of Pn−1 . Hence, applying Remark 3.4 we deduce the
following result.
Corollary 3.5. The polynomials φ1 , . . . , φn form a regular sequence in Q[X].
Next we provide a consistent condition which assures that the polynomials
dr+1
φ1 , . . . , φr , Yr+1
, . . . , Yndn form a regular sequence in Q[X] for 1 ≤ r ≤ n.
Proposition 3.6. Fix a positive integer ρ and suppose that the coefficients
of the linear forms Y1 , . . . , Yn are randomly chosen in the set {1, . . . , 2nρD}.
69
d
r+1
Then with error probability at most 1/ρ, the polynomials φ1 , . . . , φr , Yr+1
,...,
dn
Yn form a regular sequence in Q[X] for 0 ≤ r ≤ n.
Proof. For 1 ≤ r ≤ n, we deduce from Corollary 3.5 that the affine variety
Wr,0 := {x ∈ An ; φ1 (x) = 0, . . . , φr (x) = 0}
is equidimensional of dimension n − r for 1 ≤ r < n. This in particular shows
the condition of the statement of the proposition for r = n is automatically
satisfied.
Let Ai,j (1 ≤ i, j ≤ n) be new indeterminates over Q[X], and denote
A(r) := (Ar,1 , . . . , Ar,n ) for 1 ≤ r ≤ n and A := (Ai,j )1≤i,j≤n .
We construct the linear forms Y1 , . . . , Yn iteratively so that at step r:
the coefficients of the linear forms Y1 , . . . , Yr are randomly chosen in
the set {1, . . . , 2nρD}, and
for any pair (i, j) with 0 ≤ i < j ≤ r, the variety
Wi,j := {x ∈ An ; φ1 (x) = 0, . . . , φi (x) = 0, Yi+1 (x) = 0, . . . , Yj (x) = 0}
(3.5)
is
equidimensional
of
dimension
n
−
j
with
error
probability
at
most
Pr
Pj−1
j=1 (1 +
i=1 d1 · · · di )/2nρD .
Fix arbitrarily a nonzero linear form Y1 ∈ Q[X] with coefficients in
{1, . . . , 2nρD} and an index r with 1 < r ≤ n. If r = 1, then the conditions are satisfied, since for any nonzero linear form Y1 the affine variety
W0,1 is equidimensional of dimension 1.
Assume r > 1 and that Y1 , . . . , Yr−1 have been chosen fulfilling the conditions above. We can now discuss the choice of the linear form Yr . For any
i with 0 ≤ i ≤ r − 1, let Si,r−1 ⊂ Wi,r−1 be a finite set consisting of one arbitrary nonzero point in each irreducible component of Wi,r−1 . By the Bézout
inequality (2.1) we have #Si,r−1 ≤ deg Wi,r−1 ≤ d1 . . . di for i = 1, . . . , r − 1
and #S0,r−1 = deg W0,r−1 = 1. Let Qr ∈ Q[A(r) ] be the nonzero polynomial
defined in the following way:
(r)
Qr (A ) :=
n
r−1 Y X
Y
i=0 ξ∈Si,r−1 j=1
ξj Ar,j .
70
Let a(r) be an arbitrary point of Qn not annihilating Qr and define Yr :=
a(r) X. By construction, we have that Yr (ξ) 6= 0 holds for every ξ ∈ Si,r−1
and every 0 ≤ i ≤ r − 1. This shows that the hyperplane {Yr = 0} cuts
properly all the irreducible components of Wi,r−1 for 0 ≤ i ≤ r − 1. We
conclude that the variety
Wi,r := {x ∈ An ; φ1 (x) = 0, . . . , φi (x) = 0, Yi+1 (x) = 0, . . . , Yr (x) = 0}
(3.6)
is equidimensional of dimension n − r and degree at most d1 · · · di for 0 ≤ i ≤
r. Combining (3.5), (3.6) we see that the varieties Wi,j are equidimensional
of dimension n − j for 0 ≤ i < j ≤ r, completing thus the rth step of our
inductive argument.
Pr−1
Observe that deg Qr ≤ 1 + i=1
d1 · · · di holds. Hence, applying Theorem 2.1 we conclude that for a random choice of the coefficient vector a(r)
(r)
of Yr in the set {1, . . . , 2nρD}n , the inequality
Pr−1Qr (a ) 6= 0, and thus (3.6),
holds, with error probability at most (1 + i=0 d1 · · · di )/2nρD.
After n steps as described, we
Pnobtain linear
Pj−1 forms Y1 , . . . , Yn such that,
with error probability at most j=1 (1 + i=1 d1 · · · di )/2nρD, the variety
Wi,j is equidimensional of dimension n − j for i < j ≤ n and 0 ≤ i ≤ n. In
particular, this implies that the polynomials φ1 , . . . , φr , Yr+1 , . . . , Yn form a
regular sequence of Q[X] for 0 ≤ r ≤ n. Therefore, by [Mat86, Theorem 16.1]
dr+1
, . . . , Yndn form a regular sequence in Q[X]
we conclude that φ1 , . . . , φr , Yr+1
for 0 ≤ r ≤ n. Since
j−1
n
n
X
X
1 X
1 n+
1+
(n − j)d1 · · · dj
d1 · · · di =
2nρD j=1
2nρD
j=1
i=1
!
n
X
1
1
1+
d1 · · · dj ≤ ,
=
2ρD
ρ
j=1
we deduce the statement of the proposition.
As a first consequence on the choice of the linear forms Y1 , . . . , Yn we
deduce that for a generic value t ∈ A1 the polynomials in (3.3) define a
generalized Pham system. More precisely, we have the following result.
71
Corollary 3.7. Let Y1 , . . . , Yn ∈ Q[X] be linear forms satisfying the statement of Proposition 3.6. Then, for 1 ≤ r ≤ n, the polynomials
d
r+1
φ1 (X), . . . , φr−1 (X), Yr+1
, . . . , Yndn , Yrdr + T (φr (X) − Yrdr )
(3.7)
form a regular sequence in Q[T, X]. Furthermore, for a generic value t ∈ A1
d
r+1
φ1 (X), . . . , φr−1 (X), Yr+1
, . . . , Yndn , Yrdr + t(φr (X) − Yrdr )
form a regular sequence in C[X].
Proof. Fix r with 1 ≤ r ≤ n. From Proposition 3.6 it follows that the polydr+1
nomials φ1 , . . . , φr , Yr+1
, . . . , Yndn form a regular sequence in Q[X]. Since
these are homogeneous polynomials, from [Mat86, Corollary of Theorem
dr+1
16.3] we see that φ1 , . . . , φr−1 , Yr+1
, . . . , Yndn form a regular sequence of
Q[X]. Therefore, in order to prove the lemma it suffices to show that
Hr (T, X) := Yrdr + T (φr − ar Yrdr ) is not a zero divisor modulo the ideal
dr+1
, . . . , Yndn ) ⊂ Q[T, X].
Ir∗ := (φ1 , . . . , φr , Yr+1
Let V (Irc ) := ∪j∈J Cj be the decomposition of the projective subvariety
dr+1
of Pn−1 defined by the ideal Irc := (φ1 , . . . , φr , Yr+1
, . . . , Yndn ) ⊂ Q[X]. Fix
(j)
a point p̄ in each irreducible component Cj of V (Irc ). Since Yrdr and φr
are not zero divisors modulo Irc , it follows that αj := Yr (p̄(j) )dr and βj :=
φr (p̄(j) ) are nonzero complex numbers and hence αj + T (βj − αj ) is a nonzero
polynomial
of C[T ] for every j ∈ J . Finally, let t ∈ A1 be any value with
Q
dr
dr
j∈J (αj + t(βj − αj )) 6= 0. Then Yr + t(φr (X) − Yr ) is not a zero divisor
modulo Irc . Furthermore, taking into account that V (Ir∗ ) = ∪j∈J (A1 × Cj )
is the decomposition
of V (Ir∗ ) ⊂ Pn into irreducible components, from the
Q
condition j∈J (αj + t(βj − αj )) 6= 0 we conclude that Yrdr + T (φr − Yrdr ) is
not a zero divisor modulo Ir∗ . This finishes the proof.
d
r+1
A consequence of Corollary 3.7 is that φ1 , . . . , φr−1 , Yrdr +t(φr −Yrdr ), Yr+1
,
dn
n−1
. . . , Yn define the empty projective variety of P
for all but a finite number
of t ∈ A1 . Likewise, from Corollary 3.5 we conclude that (3.4) is a generalized
Pham system for every substitution T = t. Therefore, we have the following
result.
Corollary 3.8. Let Y1 , . . . , Yn be linear forms satisfying the statement of
Proposition 3.6. Then for all but a finite number of values t ∈ A1 , making
the substitution T = t in (3.3) and (3.4) yields a generalized Pham system
for 1 ≤ r ≤ n + 1.
72
This proves that our choice of Y1 , . . . , Yn implies that condition (A) in
Section 3.2.2 holds.
Next we consider condition (B). Fix r with 1 ≤ r ≤ n + 1 and let
b1 , . . . , bn be nonzero rational numbers to be fixed. Condition (B) asserts
(r)
(r)
that the affine subvariety V (F1 (0, X), . . . , Fn (0, X)) of An consists of D
distinct points, none of which is annihilated by the Jacobian determinant
Jr (0, X) := det(∂Fi(r) (0, X)/∂Xj )1≤i,j≤n .
We observe that the smoothness of the variety V(F1(r) (0, X), . . . , Fn(r) (0, X))
for a generic value of b is a direct consequence of the Sard’s theorem (see,
e.g., [GP74, §1.7]). The next lemma is an effective version of this result in
our context.
Proposition 3.9. Let B1 , . . . , Bn be new indeterminates over Q[X] and set
B := (B1 , . . . , Bn ). Let Y1 , . . . , Yn ∈ Q[X] be linear forms satisfying the
statement of Proposition 3.6. Then there exists a nonzero polynomial P (2) ∈
Q[B] of degree at most 2nD2 such that for every b ∈ Qn with P (2) (b) 6= 0, the
(r)
(r)
affine variety defined by the polynomials F1 (0, X), . . . , Fn (0, X) for this
value of b consists of D nonsingular points for 1 ≤ r ≤ n + 1.
Proof. First we observe that choosing arbitrarily b1 6= 0, . . . , bn 6= 0, the
affine variety {Y1 (x)d1 + b1 = 0, . . . , Yn (x)dn + bn = 0} consists of D
distinct points which are not annihilated by the Jacobian J1 (0, X) :=
det(∂Yi (x)di /∂Xk )1≤i,j≤n , since no point of this variety has a zero coordinate. This proves the result for r = 1.
Now fix r with 2 ≤ r ≤ n and consider the following polynomials of
Q[B, X]:
φ1 (X) + B1 , . . . , φr−1 (X) + Br−1 , Yrdr + Br , . . . , Yndn + Bn .
(3.8)
Denote by Wr the affine subvariety of A2n defined by these polynomials.
Since the polynomials in (3.8) define a generalized Pham system for any
substitution B = b, from Lemma 3.1 it follows that the morphism
θr : W r → An ,
θr (b, x) := b
is surjective. Furthermore, we claim that θr is finite. Indeed, let Nr ∈ N with
(r)
Nr > d and hi,j ∈ Q[X] (1 ≤ i, j ≤ n) be polynomials of degree Nr − dr such
that
XiNr
=
r−1
X
(r)
hi,j φj
+
j=1
Then for every 1 ≤ i ≤ n we have
r−1
X
j=1
(r)
hi,j (φj
+ Bj ) +
n
X
j=r
d
(r)
hi,j (Yj j
n
X
(r)
73
d
hi,j Yj j .
j=r
+ Bj ) =
XiNr
+
n
X
(r)
hi,j Bj ,
j=1
(r)
with deg hi,j Bj < Nr for 1 ≤ i, j ≤ n. This proves that Q[B] ֒→ Q[Wr ] is
an integral ring extension. Taking into account that Wr is irreducible, from
the previous assertions and [Sha94, II.6.3, Theorem 4] we conclude that θr is
generically unramified.
Let η ∈ C[X] be a linear form that induces a primitive element of the ring
(r)
extension C[B] ֒→ C[Wr ]. Consider its minimal polynomial Mη ∈ Q[B, Y ],
and let ∆r ∈ C[B] denote its discriminant with respect to the variable Y . For
every b ∈ Qn such that ∆r (b) 6= 0 it follows that the Mη (b, Y ) is square–free,
and therefore the morphism θr is unramified at B = b.
Furthermore, for every b ∈ Qn with ∆r (b) 6= 0, the corresponding polynomials φ1 (X)+b1 , . . . , φr−1 (X)+br−1 , Yrdr +br , . . . , Yndn +bn define a generalized
h
h
Pham system and generate a radical ideal in Q[X]. Let fr,1
, . . . , fr,n
denote
dr
the homogenizations of φ1 (X) + b1 , . . . , φr−1 (X) + br−1 , Yr + br , . . . , Yndn +
h
h
bn with homogenizing variable X0 . Then fr,1
, . . . , fr,n
is a radical zero–
h
h
dimensional ideal. We conclude that fr,1 , . . . , fr,n form a regular sequence
of Q[X]. Applying the Bézout theorem in the form of [EH99, Theorem
h
h
III.71] we see that the projective variety defined by fr,1
, . . . , fn,r
has preh
n
cisely deg V = D distinct points in P . Finally, since there are no points at
infinity, from [CGH91, Proposition 1.11] we conclude that θr−1 (b) consists of
exactly D distinct points.
A similar argument shows that there exists a polynomial ∆n+1 ∈ Q[B]
(n+1)
such that, for every b ∈ Qn with ∆n+1 (b) 6= 0, the polynomials F1
(0, X),
(n+1)
(0, X) corresponding to this value of b have D nonsingular common
. . . , Fn
zeros.
Let P (2) := (B1 · · · Bn ) · ∆2 · · · ∆n+1 . From the Bézout inequality (2.1)
we deduce that deg Wr ≤ D holds, which in turns implies that the minimal
polynomial of η has degree bounded by D. It follows that the degree of its
74
discriminant ∆r is bounded by 2D2 for 2 ≤ r ≤ n + 1 and hence deg P (2) ≤
n + 2(n − 1)D2 ≤ 2nD2 holds. In conclusion, the polynomial P (2) satisfies
all the requirement of the proposition.
Fix a positive integer ρ. From Theorem 2.1 it follows that for a random
choice of b1 , . . . , bn in the set {1, . . . , 2nρD2 }, the inequality P (2) (b1 , . . . , bn ) 6=
0 holds with error probability at most 1/ρ. Then we have the following result.
Corollary 3.10. If b1 , . . . , bn are randomly chosen in the set {1, . . . , 2nρD2 },
(r)
(r)
then the variety defined by F1 (0, X), . . . , Fn (0, X) for such values b1 , . . . , bn
consists of D nonsingular points for 1 ≤ r ≤ n + 1 with error probability at
most 1/ρ.
3.2.4.
The main algorithm
By means of Propositions 3.6 and 3.9 we obtain deformations πr : Vr →
A satisfying conditions (i)–(v) of Section 3.2.1. In this section we present a
more detailed outline of Algorithm 3.2 and estimate its complexity and error
probability.
1
Algorithmic tools
In order to present a more detailed outline of Algorithm 3.2, we need to
make explicit the procedures that arise during the execution of this algorithm:
the procedure for computing a geometric solution of a zero–dimensional
diagonal square system used in the first step,
the “projection algorithm” used in the second step,
the procedure used to clean multiplicities used in the third step.
For the last two procedures, we shall use the algorithms underlying Theorem 2.2 and Lemma 2.3. A procedure for solving a given zero–dimensional
diagonal square system follows.
75
Lemma 3.11. Let be given Q–linearly independent linear forms Y1 , . . . , Yn ∈
Q[X] and nonzero rational numbers b1 , . . . , bn . Set gi (X) := Yi (X)di + bi
(1 ≤ i ≤ n) and let V0 ⊂ An be the affine variety defined by g1 , . . . , gn .
Then, given a generic linear form u ∈ Q[X] we can compute a geometric
solution of V0 with O(nM(D2 )) arithmetic operations in Q.
Proof. Suppose that we are given a linear form u := λ1 X1 +· · ·+λn Xn ∈ Q[X]
that induces a primitive element of the Q–algebra extension Q → Q[V0 ]. We
can compute the minimal polynomial of u as follows: let Y, Z be new variables
and set
m1 (Y ) := λ1−d1 Y d1 − b1 ,
dr
r
mr (Y ) := ResZ λ−d
r (Y − Z) − br , mr−1 (Z) (2 ≤ r ≤ n).
(3.9)
We claim that the polynomial mn equals (up to scaling by a nonzero element
of Q) the minimal polynomial mu ∈ Q[Y ] of u in Q → Q[V0 ]. Indeed, for
r
every 2 ≤ r ≤ n, the polynomial mr (Y ) is a linear combination of λ−d
r (Y −
dr
Z) − br and mr−1 (Z) over Q[Y, Z] by properties of the resultant. Let
u(r) := λ1 X1 + · · · + λr Xr for 1 ≤ r ≤ n. Then, the identity
(r)
r
− u(r−1) )dr − br = 0
λ−d
r (u
holds in Q[V0 ]. Thus, assuming inductively that mr−1 (u(r−1) ) = 0 holds in
Q[V0 ], it follows that mr (u(r) ) = 0 holds in Q[V0 ] as well. Since m1 (u(1) ) =
X1d1 −b1 = 0 in Q[V0 ] it holds that mn (u(n) ) = 0 in Q[V0 ]. Taking into account
the estimate deg mn ≤ D and the fact that mu is a nonzero polynomial of
degree D = #(V0 ), we conclude that our claim holds.
In order to compute the polynomial mu , we compute the
resultants in
dr
r
(Y
−
Z)
−
b
,
m
(Z)
is a polynomial
(3.9). Since the resultant ResZ λ−d
r
i−1
r
of Q[Y ] of degree d1 · · · dr , by univariate interpolation in the variable Y we
reduce its computation to the computation of d1 · · · dr + 1 resultants of uni-
variate polynomials in Q[Z]. This interpolation step requires O M(d21 · · · d2r )
arithmetic operations in Q and does not require any division by a nonconstant polynomial in the coefficients λ1 , . . . , λn (see, e.g., [BLS03], [BS05]).
Each univariate resultant can be computed with M(d1 · · · dr ) arithmetic operations in Q (see Section 2.3.3). Altogether,
we obtain an algorithm for
2
computing mu which performs O M(D ) arithmetic operations in Q.
76
Finally, by the genericity of the linear form u we see that we can extend
this algorithm to an algorithm computing a geometric solution of V0 as explained in Section 2.4.2. From Lemma 2.4 we deduce the statement of the
lemma.
We remark that the genericity condition underlying the choice of the
linear form u shall be discussed below.
Outline of the main algorithm and complexity and error probability
estimate
Now we can give a more detailed outline of the algorithm computing a
geometric solution of the input variety V .
Algorithm 3.12 (Algorithm for solving f1 = 0, . . . , fn = 0).
1. Choose linear forms Y1 , . . . , Yn ∈ Q[X] and b1 , . . . , bn ∈ Q \ {0} according to Propositions 3.6 and 3.9. (These choices determine the deformations π1 , . . . , πn+1 .)
2. Choose randomly a linear form u ∈ Q[X] which induces a primitive
element of V and πr−1 (0) for 1 ≤ r ≤ n + 1 and such that we can apply
Lemma 3.11.
3. Since π1−1 (0) = {Yi (x)di + bi = 0; 1 ≤ i ≤ n} holds, then the fibre
π1−1 (0) is defined by a diagonal system. Apply the algorithm underlying
Lemma 3.11 in order to compute a geometric solution of π1−1 (0) with u
as primitive element.
4. For r = 1 to n + 1 do:
a) Use the “projection algorithm” underlying Lemma 2.2 in order to
compute a geometric solution of Vr with u as primitive element,
from the geometric solution of πr−1 (0) computed in the previous
step.
−1
−1
b) The equalities πr−1 (1) = πr+1
(0) (1 ≤ r ≤ n) and πn+1
(1) =
{1} × V hold. Make the substitution T = 1 in the polynomials that
form the geometric solution of Vr computed in the previous step.
The univariate polynomials obtained form a geometric solution
77
of πr−1 (1) for 1 ≤ r ≤ n and a complete description (eventually
−1
including multiplicities) of πn+1
(1) = {1} × V for r = n + 1.
5. Apply the algorithm underlying Lemma 2.3 to the polynomials that form
−1
a complete description of πn+1
(1) = {1} × V computed in the previous
step. The output is a geometric solution of V .
In order to estimate the cost of this algorithm we need to estimate the
cost of computing the geometric solution of the variety Vr of the step 4(a) for
each r with 2 ≤ r ≤ n + 1. According to Lemma 2.2 such a cost depends on
the height Er of the projection πr : Vr → A1 , namely, the degree degT Mu(r)
in T of the minimal polynomial Mu(r) ∈ Q[T, Y ] of a generic linear form u ∈
Q[X1 , . . . , Xn ] in Vr . While the obvious estimate En+1 ≤ D is an immediate
consequence of the Bézout inequality (2.1), in order to estimate Er for 1 ≤
r ≤ n, we have the following result (cf. Lemma 5.3).
Lemma 3.13. The inequality Er ≤ D/dr holds for 2 ≤ r ≤ n.
(r)
Proof. Observe that substituting a generic value y ∈ Q for Y in Mu (T, Y )
we have degT Mu(r) (T, Y ) = degT Mu(r) (T, y) = #{t ∈ C; Mu(r) (t, y) = 0}.
Moreover, it follows that Mu(r) (t, y) = 0 if and only if there exists x ∈ An
with y = u(x) and (t, x) ∈ Vr . Thus, it suffices to estimate the number of
(r)
(r)
points (t, x) ∈ An+1 with u(x) − y = 0, F1 (t, x) = 0, . . . , Fn (t, x) = 0.
Being u generic, the system
(r)
u(X) − y = 0, F1 (T, X) = 0, . . . , Fn(r) (T, X) = 0
(r)
(3.10)
(r)
(r)
has finitely many solutions in An+1 . Furthermore, u(X)−y, F1 , . . . , Fr−1 , Fr+1
(r)
, . . . , Fn ∈ Q[X] define a zero–dimensional variety Wr of An of degree
at most D/dr . Let ℓ ∈ Q[X] be a separating linear form of Wr , and let
mℓ , w1 , . . . , wn ∈ Q[Y ] be a geometric solution of Wr . Then an eliminating
(r)
(r)
polynomial for T modulo (u − y, F1 , . . . , Fn ) is the resultant qT (Y ) :=
(r)
ResY mℓ (Y ), Fr (T, w(Y )) . It is easy to see that degT qT ≤ D/dr , which
implies the statement of the lemma.
The following result is a critical step for our error probability estimate.
Proposition 3.14. Suppose that the coefficients of the linear form u are
randomly chosen in the set {1, . . . , 4nρD3 }, where ρ is a fixed positive integer.
Then the following assertions hold with error probability at most 1/ρ:
78
u separates the points of V and the fibres πr−1 (0) for 1 ≤ r ≤ n + 1,
the algorithm underlying Lemma 3.11 outputs the right result.
Proof. Let Λ1 , . . . , Λn be new indeterminates and set UΛ := Λ1 X1 + · · · +
Λn Xn . Fix r with 1 ≤ r ≤ n + 1. From Proposition 3.9 we see that πr−1 (0) is
a zero–dimensional variety of degree D. Let πr−1 (0) := {ξ (1) , . . . , ξ (D) }. Then
the nonzero polynomial
Y
P (r) :=
UΛ (ξ (i) ) − UΛ (ξ (j) )
1≤i<j≤D
has degree D(D − 1)/2 and satisfies the following condition: for any λ ∈ Qn
with P (r) (λ) 6= 0, the linear form u := λ1 X1 +· · ·+λn Xn separates the points
of πr−1 (0). Similarly, let V := {ζ (1) , . . . , ζ (δ) } and set
Y
Q :=
UΛ (ζ (i) ) − UΛ (ζ (j) ) .
1≤i<j≤δ
Finally, define P := P (1) · · · P (n+1) Q and observe that deg P = (n + 1)D(D −
1)/2 + δ(δ − 1)/2 ≤ nD2 . For any λ ∈ Qn with P (λ) 6= 0, the linear form
u := λ1 X1 + · · · + λn Xn separates the points of V and the fibres πr−1 (0) for
1 ≤ r ≤ n + 1.
From Theorem 2.1 it follows that for a random choice of the coefficients of
u in the set {1, . . . , 4nρD3 }, the linear form u separates the points of πr−1 (0)
for 1 ≤ r ≤ n + 1 and V with error probability at most 1/4ρD ≤ 1/2ρ.
Next we consider the second requirement of the proposition, namely, the
computation of the univariate resultants of the generic versions of the polynomials in (3.9). This is required in order to extend the algorithm for computing the minimal polynomial of u in π1−1 (0) to an algorithm for computing
a geometric solution of π1−1 (0). We use a fast algorithm for computing resultants over Q(Λ) based on the EEA (Extended Euclidean Algorithm; see
Section 2.3.3). We perform the EEA over the ring of power series Q[[Λ − λ]],
truncating all the intermediate results up to order 2. Therefore, the choice
of the coefficients of u must guarantee that all the elements of Q[Λ] which
have to be inverted during the execution of the EEA are invertible elements
of Q[[Λ − λ]].
For this purpose, we observe that, similarly to the proof of [vzGG99,
Theorem 6.52], one deduces that all the denominators of the elements of
79
Q(Λ) arising during the application of the EEA to the generic version of the
(r−1) dr
r
polynomials λ−d
) − br and mr−1 (u(r−1) ) are divisors of at most
r (α − u
d1 · · · dr−1 polynomials of Q[Λ] of degree 2d1 · · · dr for any α ∈ Q. This EEA
step must be executed for 1 + d1 · · · dr distinct values of α ∈ Q, in order
to perform the interpolation step. Hence the product of the denominators
arising during all the applications of the EEA has degree at most 2nD3 .
Therefore, from Theorem 2.1 we conclude that for a random choice of its
coefficients in the set {1, . . . , 4nρD3 }, the linear form u satisfies our second
requirement with error probability at most 1/2ρ.
Adding both probability estimates finishes the proof of the proposition.
Now we can estimate the complexity and error probability of Algorithm 3.12.
Theorem 3.15. Suppose that we are given a division–free straight–line program of length T evaluating polynomials φ1 , . . . , φn , ϕ1 , . . . , ϕn ∈ Q[X] such
that φi is homogeneous of degree di > 0 and deg ϕi < di holds for 1 ≤ i ≤ n.
Assume that f1 := φ1 + ϕ1 , . . . , fn := φn + ϕn define a generalized Pham
system. Furthermore, fix a positive integer ρ and suppose that, for D :=
d1 · · · dn ,
the coefficients of the linear forms Y1 , . . . , Yn of step (1) of Algorithm
3.12 are randomly chosen in the set {1, . . . , 6nρD},
the rational numbers b1 , . . . , bn of step (1) of Algorithm 3.12 are randomly chosen in the set {1, . . . , 6nρD2 },
the coefficients of the linear form u of step (2) of Algorithm 3.12 are
randomly chosen in the set {1, . . . , 12nρD3 }.
Then Algorithm 3.12 computes a geometric solution of V with
P
O (nT + n3 )M(D) n+1
r=1 M(D/dr )
(3.11)
arithmetic operations in Q and error probability at most 1/ρ, where dn+1 := 1.
Proof. According to Proposition 3.6, Corollary 3.10 and Proposition 3.14,
for a random choice of the coefficients of Y1 , . . . , Yn , u and b1 , . . . , bn as in
the statement of the theorem, with probability at least 1/ρ the following
assertions hold:
80
(a) the linear forms Y1 , . . . , Yn satisfy the statement of Proposition 3.6,
(b) the rational numbers b1 , . . . , bn satisfy the statement of Corollary 3.10,
(c) the linear form u satisfy the statement of Proposition 3.14.
From (c) we conclude that the algorithm underlying
Lemma 3.11 com
−1
2
putes a geometric solution of π1 (0) with O nM(D) arithmetic operations
in Q.
From (b) and (c) we see that the deformations πr : Vr → A1 satisfy all
the requirements of Lemma 2.2 for 1 ≤ r ≤ n + 1. Furthermore, from the
input straight–line program evaluating φ1 , . . . , φn , ϕ1 , . . . , ϕn we may easily
(r)
obtain a straight–line program evaluating the polynomials Fi (1 ≤ i ≤ n)
of (3.3) or (3.4) with T + O(n) arithmetic operations in Q. Hence, applying
the algorithm underlying Lemma 2.2 successively to the fibre πr−1 (0) and the
variety Vr for r = 1, . . . , n + 1, we finally obtain polynomials mu , v1 , . . . , vn ∈
Q[T, Y ] which form a geometric solution of Vn+1 . These steps require O (nT+
P
n3 )M(D) n+1
M(D/d
)
arithmetic operations in Q.
r
r=1
Finally, we apply the algorithm underlying Lemma 2.3 to the polynomials
mu , v1 , . . . , vn ∈ Q[T, Y ] and obtain a geometric solution of V with O(nM(D))
additional arithmetic operations in Q. This finishes the proof of the theorem.
Chapter 4
Polynomial equation solving by
lifting procedures for ramified
fibres
Let V be a Q–definable equidimensional affine variety of dimension 1, a
curve, and let be given a generically unramified, finite morphism π : V → C.
Earlier in this thesis, we have exhibited an algorithm which computes a
geometric solution of V given a geometric solution of a particular unramified
fibre π −1 (ε0 ) by using a global version of the Newton–Hensel lifting. Now
we introduce an algorithm which, given a generically unramified family of
zero–dimensional affine varieties, represented by a dominant (not necessarily
finite) morphism π : V → C, and the infinitesimal structure of a particular
(eventually ramified) fibre π −1 (ε0 ), computes a complete geometric solution
of V .
This chapter is based on an article with the same title which I co-authored
with A. Bompadre, G. Matera and R. Wachenchauzer ([BMWW04]). More
explicitly, let V ⊂ Cn+1 be a Q–definable algebraic curve, and let us assume
that the morphism π : V → C induced by the canonical projection in the
first coordinate is dominant and generically unramified. Let π −1 (ε0 ) be a
finite and ramified fibre. Suppose further that we are given the infinitesimal
structure of π −1 (ε0 ), i.e. the set of singular parts of the Puiseux expansions
of the branches of V lying above ε0 (see Section 4.1 for further details).
Then we exhibit an algorithm which computes a complete description of an
81
82
CHAPTER 4. DEFORMATIONS FOR RAMIFIED FIBRES
arbitrary fibre π −1 (ε) (see Section 4.3).
Our algorithmic method is essentially based on a new variant of the global
Newton–Hensel procedure. It is described in Section 4.2. Its time–space
complexity is roughly O(deg V (deg π)α ), where α = 1 in several important
cases. Then our algorithm extends and improves the procedures in [HKP+ 00]
and [Sch03]. Furthermore, our algorithm treats all the branches of V lying
above ε0 separately, improving thus the refinements of [HKP+ 00, Section 3].
4.1.
Puiseux expansions of space curves
In this section we introduce terminology about space curves, extending
the usual terminology of Puiseux expansions of plane curves (see e.g. [Wal50])
and rational Puiseux expansions (see e.g. [Duv89], [Wal99]).
Let T, E, X1 , . . . , Xn be indeterminates over Q. Let n be a fixed positive
integer, and An and An+1 denote the affine spaces An (C) and An+1 (C). We
denote their coordinates by x ∈ Cn and (ε, x) ∈ Cn+1 with ε ∈ C.
Let F1 , . . . , Fn be polynomials in Q[E, X] = Q[E, X1 , . . . , Xn ] which form
a regular sequence and generate a radical ideal in Q[E, X]. Let V := {(ε, x) ∈
An+1 : F1 (ε, x) = 0, . . . , Fn (ε, x) = 0}, and note that this is a curve.
Let π : V → A1 be the morphism induced by the restriction to V of
the canonical projection in the first coordinate π(ε, x) := ε. Assume that π
is generically unramified; this implies that the Jacobian determinant JF :=
det(∂Fi /∂Xj )1≤i,j≤n is not a zero divisor in Q[V ].
e X)
e of eleA parameterization of the curve V is a non–constant vector (E,
e
e
e
ments of the field of Laurent series Q((T )), with X := (X1 , . . . , Xn ) ∈ Q((T ))n ,
e X)
e = 0, . . . , Fn (E,
e X)
e = 0 holds in Q((T )). A parameterizasuch that F1 (E,
e X)
e is called irreducible if there does not exist an integer k > 1 for
tion (E,
e X)
e ∈ Q((T k ))n+1 holds. The coefficient field of a parameterization
which (E,
e X)
e of V is the field extension of Q generated by the coefficients of the
(E,
eX
e1 , . . . , X
en .
series E,
Given an element ϕ ∈ Q((T )), we define its order oT (ϕ) in T as the least
power of T appearing with a nonzero coefficient in ϕ. Two parameterizations
e X)
e and (Ee′ , X
e ′ ) are called equivalent if there exists a power series ϕ ∈ C[[T ]]
(E,
83
4.1. PUISEUX EXPANSIONS OF SPACE CURVES
e ) = Ee′ (ϕ(T )), X
e1 (T ) = X
e ′ (ϕ(T )), . . . , X
en (T ) =
of order 1 such that E(T
1
′
e (ϕ(T )) holds in Q((T )). A branch C of the curve V is defined as the
X
n
equivalence class of an irreducible parameterization of V . We say that a
e X)
e in
branch C lies above a point ε ∈ A1 if there exists a parameterization (E,
e = ε.
the equivalence class that defines the branch C with Ee ∈ Q[[T ]] and E(0)
In what follows we shall consider the branches of V lying above 0. It is
well–known that if 0 is an unramified value of the morphism π : V → A1
defined by the rule π(ε, x) := ε, then all the branches of V lying above 0 have
e with X
e ∈ Q[[T ]]n (see Section 1.1).
a parameterization of the form (T, X)
Now we explain how the parameterizations of the branches of V lying
above 0 can be represented by means of Puiseux series in E. Let Q(E)∗ :=
∪q≥0 Q(E 1/q ) denote the field of Puiseux series in the variable E over Q, where
Q is the field of algebraic numbers. It is well–known that Q(E)∗ is an algebraically closed field. In fact, it is an algebraic closure of Q(E) (see e.g.
[Wal50]).
Let us consider F1 , . . . , Fn as elements of the polynomial ring Q(E)∗ [X].
Since the Q–algebra extension Q(E) ֒→ Q(V ) is finitely generated, it follows
that the affine variety {x̄ ∈ An (Q(E)∗ ) : F1 (x̄) = 0, . . . , Fn (x̄) = 0} has
dimension zero. Therefore, under our hypotheses there exist D := deg π
(ℓ)
(ℓ)
distinct n–tuples x(ℓ) := (x1 , . . . , xn ) ∈ (Q(E)∗ )n of Puiseux series which
are solutions of the system defined by F1 , . . . , Fn over Q(E)∗ , i.e. such that
the following equalities hold in Q(E)∗ for 1 ≤ ℓ ≤ D:
F1 (E, x(ℓ) ) = 0 , . . . , Fn (E, x(ℓ) ) = 0.
(ℓ)
(ℓ)
(ℓ)
For 1 ≤ ℓ ≤ D, let us write x(ℓ) := (x1 , . . . , xn ) and xi :=
(ℓ)
(4.1)
P
(ℓ)
m≥mℓ
m
ai,m ·E eℓ
(1 ≤ i ≤ n), with eℓ ∈ N, mℓ ∈ Z and ai,m ∈ Q. Without loss of generality we
may assume for 1 ≤ ℓ ≤ D that eℓ has no common factors with the greatest
(ℓ)
common divisor of the set of m’s for which ai,m 6= 0 holds. The number
eℓ is called the ramification index of the series x(ℓ) . Let us remark that for
1 ≤ ℓ ≤ D the coefficient field generated by all the coordinates of x(ℓ) is a
finite extension of Q (see e.g. [Duv89]). Its degree fℓ is called the residual
degree of x(ℓ) .
Following [Duv89] (see also [Wal99]), a set of non–equivalent parameterizations
e (1) ), . . . , (Ee(bg ) , X
e (bg ) )} ⊂ Q((T ))n+1
{(Ee(1) , X
(4.2)
84
containing a complete set of representatives of the branches of V lying above
0 is called a system of rational Puiseux expansions (of the branches of V lying
above 0) if it is invariant under the action of the Galois group of the field
extension Q/Q and Ee(ℓ) = λℓ T eℓ , with eℓ ∈ N and λℓ ∈ Q \ {0} for 1 ≤ ℓ ≤ gb.
Let g be the number of orbits defined on the set (4.2) under the action of
the Galois group of Q/Q and suppose that we have chosen the numbering in
(4.2) such that the first g elements represent different orbits.
Let us observe that from a given system of rational Puiseux expansions we
may easily obtain the system of classical Puiseux expansions of the branches
of V lying above 0, i.e. the complete set of solutions of (4.1). Indeed, let
o
n
X
X (ℓ)
m
e (ℓ) ) := λℓ T eℓ ,
:
1
≤
ℓ
≤
g
(4.3)
a(ℓ)
T
a1,m T m , . . . ,
(Ee(ℓ) , X
n,m
m≥mℓ
m≥mℓ
−1/e
be a system of rational Puiseux expansions of V , and let ξℓ , λℓ ℓ ∈ Q denote
a primitive eℓ –th root of 1 and an eℓ –th root of λ−1
ℓ for 1 ≤ ℓ ≤ g. Then the
classical Puiseux expansions of the branches of V lying above 0 are given by
o
n
(ℓ) j −1/eℓ 1/eℓ
e
X (ξℓ λℓ
E ) : 1 ≤ ℓ ≤ g, 1 ≤ j ≤ eℓ .
e (ℓ) (ξ j λ−1/eℓ E 1/eℓ ) is
Observe that the ramification index of the expansion X
ℓ ℓ
eP
the least integer such that the partial expansion vectors
ℓ . Let R denote P
(ℓ) m
(ℓ)
(ℓ)
R
m
:= R
are pairwise distinct for 1 ≤ ℓ ≤
m=mℓ am T
m=mℓ (a1,m , . . . , an,m )T
D.
Let us remark that a combination of [Sch03, Proposition 1] and [Duv89,
Lemma 2] yields the estimate R − mℓ ≤ 2(eℓ fℓ )2 . The integer R is called
the regularity index of the system (4.3). For 1 ≤ ℓ ≤ g, the partial expansion
PR
(ℓ) m
e (ℓ) .
is called the singular part of X
m=mℓ am T
4.2.
Lifting procedures for ramified fibres
e (ℓ) ) : 1 ≤
With notations and assumptions as in Section 4.1, let {(Ee(ℓ) , X
ℓ ≤ g} be a set of parameterizations which induces a system of rational
Puiseux expansions of the branches of V lying above 0 by the action of the
Galois group of Q/Q. For 1 ≤ ℓ ≤ g, let eℓ , fℓ ∈ N denote the ramification
85
4.2. LIFTING PROCEDURES FOR RAMIFIED FIBRES
index and the residual degree of the Puiseux expansions associated to the
parameterization
X
m
e (ℓ) ) := λℓ T eℓ ,
(Ee(ℓ) , X
a(ℓ)
T
,
(4.4)
m
m≥mℓ
P
n
(ℓ)
with am ∈ Q for 1 ≤ ℓ ≤ g, m ≥ mℓ . We have gℓ=1 eℓ fℓ = D (see, e.g.,
[Duv89]). Let R ∈ Z be the regularity index of the system of rational Puiseux
expansions (4.4). Let us recall the estimate R − mℓ ≤ 2(eℓ fℓ )2 on the size of
the singular parts of the parameterizations in (4.4) from Section 4.1.
Let T, Y1 , . . . , Yn be indeterminates over Q and write Y := (Y1 , . . . , Yn ).
(ℓ)
(ℓ)
Let K (ℓ) := Q({λℓ , am,1 , . . . , am,n : m ≥ mℓ }) be the coefficient field of the
e (ℓ) ). Denote by σ1(ℓ) , . . . , σ (ℓ) the morphisms of the
parameterization (Ee(ℓ) , X
fℓ
Galois group of the field extension Q ֒→ K (ℓ) . For any (ℓ, j, k) ∈ N3 with
(ℓ,k)
1 ≤ ℓ ≤ g, 1 ≤ j ≤ n and 1 ≤ k ≤ fℓ , let us define Gj ∈ Q[T, Y ] by:
(ℓ,k)
Gj
:= T
αj,ℓ
Fj
(ℓ)
σk (λℓ )T eℓ ,
R−1
X
(ℓ) (ℓ)
)T m
σk (am
m=mℓ
(ℓ,k)
where αj,ℓ ∈ Z is chosen such that the order of Gj
+YT
R
,
(4.5)
in T equals zero.
Our algorithmic methods are based on a deformation technique which
allows us to compute an arbitrary fibre of the morphism π : V → A1 by
“lifting” the fibre π −1 (0). In order to perform this process of lifting, we
would like to use a global Newton–Hensel procedure as in Chapter 1 (see
also [GHM+ 98], [GHH+ 97], [HKP+ 00], [Sch03]). Unfortunately, this is no
longer possible because the essential hypothesis on the unramifiedness of the
fibre π −1 (0) is missed.
In order to circumvent this difficulty, one might try to proceed as in the
plane curve case and consider the ideal I (ℓ,k) of Q[T, Y ] generated by the
(ℓ,k)
(ℓ,k)
polynomials G1 , . . . , Gn for 1 ≤ ℓ ≤ g and 1 ≤ k ≤ fℓ . Let V (ℓ,k)
be the affine sub-variety of An+1 defined by I (ℓ,k) , and let π (ℓ,k) : V (ℓ,k) →
A1 be the morphism defined by π (ℓ,k) (t, x) := t. Unlike the plane curve
(ℓ,k)
(ℓ,k)
case, G1 , . . . , Gn are not necessarily smooth at T = 0, unless a suitable
flatness condition is satisfied (compare [BM93], [ANMR91] and [ANMR92]).
In Section 4.2.2 we exhibit a flatness condition which assures that the points
(ℓ,k)
(ℓ,k)
of the fibre (π (ℓ,k) )−1 (0) are (G1 , . . . , Gn )–smooth. Then in Section 4.2.3
86
we describe a variant of the global Newton–Hensel procedure of Section 2.4.1
specifically adapted to our situation. Finally, in Section 4.2.4 we show that
this flatness condition is also necessary to assure smoothness.
Let us observe that the main results of this section, namely Theorems 4.5
and 4.7 below, depend on the infinitesimal structure of the fibre π −1 (0), and
hence can be (slightly) generalized to the case where F1 , . . . , Fn form a regular
sequence of Q[E](E) [X] and generate a radical ideal of Q[E](E) [X]. Nevertheless, for the sake of clarity we are not going to prove this generalization.
4.2.1.
Properties of the ideal I (ℓ,k)
Let us fix integers ℓ, k with 1 ≤ ℓ ≤ g and 1 ≤ k ≤ fℓ . In order to exhibit
our flatness condition we first need to establish some properties of the ideal
I (ℓ,k) .
Let I (ℓ,k) Q(T )∗ [Y ] denote the (extended) ideal generated by I (ℓ,k) in
Q(T )∗ [Y ]. In order to describe the zero set of I (ℓ,k) Q(T )∗ [Y ] in An (Q(T )∗ ),
for any pair (ℓ, k), let Lℓ,k be the set of pairs (ℓ ′ , k ′ ) for which there exists a
vector of Puiseux series associated to the (ℓ ′ , k ′ )–th parameterization which
agrees up to order R with one associated to the (ℓ, k)–th parameterization,
i.e.
n
−1/e −1/e Lℓ,k := (ℓ ′ , k ′ ); eℓ = eℓ ′ , mℓ = mℓ ′ , ∃ λℓ ℓ ∃ λℓ ′ ℓ ′
R−1
R−1
o
1
X
X
− e1
(ℓ ′ )
(ℓ ′ ) (ℓ ′ )
(ℓ) − eℓ m m
(ℓ)
m
ℓ′ m
.
)
T
)
T
=
(λ
(a
)σ
σ
(λ
)σ
σk (a(ℓ)
m
m
ℓ
k
ℓ′
k′
k′
m=mℓ
m=mℓ
(4.6)
The sets Lℓ,k form a partition of the set of pairs ∪1≤ℓ≤g {ℓ} × {1, . . . , fℓ }. For
any (ℓ, k) let
Ve (ℓ,k) := V I (ℓ,k) Q(T )∗ [Y ]
be the affine subvariety of An (Q(T )∗ ) induced by I (ℓ,k) Q(T )∗ [Y ].
Lemma 4.1. The extended ideal I (ℓ,k) Q(T )∗ [Y ] defines a zero–dimensional
subvariety Ve (ℓ,k) of An (Q(T )∗ ). Furthermore, we have
(
)
1
X (ℓ ′ ) ′ (ℓ ′ ) − e1 ′
eℓ m
)
m−R
ℓ m (ℓ)
σ ′ (a(ℓ
; (ℓ′ ,k ′ ) ∈ Lℓ,k .
Ve (ℓ,k) ∩Q[[T ]]n=
m )σ ′ (λ ′ ) σ (λ ) T
k
k
ℓ
k
ℓ
m≥R
(4.7)
87
(ℓ,k)
(ℓ,k)
and the parameterization
Proof. From the definition of G1 , . . . , Gn
(ℓ) (ℓ)
m−R
(ℓ)
(ℓ)
e ), it follows that the vector of power series P
(Ee , X
m≥R σk (am )T
is a point of Ve (ℓ,k) ⊂ An (Q(T )∗ ).
On the other hand, we observe that any point of Ve (ℓ,k) induces univocally
a finite set of points x ∈ An (Q(E)∗ ) such that Fj (E, x) = 0 holds in Q(E)∗ for
1 ≤ j ≤ n. Since {x ∈ An (Q(E)∗ ) : F1 (x) = 0, . . . , Fn (x) = 0} has dimension
zero (see Subsection 4.1), it follows that Ve (ℓ,k) must also have dimension zero.
Now we show identity (4.7). Let Vb (ℓ,k) be the right–hand side of identity
(4.7):
Vb (ℓ,k) :=
nX
(ℓ ′ )
− e1
(ℓ ′ )
′
)
σk ′ (a(ℓ
m )σk ′ (λℓ ′
m≥R
ℓ′
o
1
e
(ℓ)
)m σk (λℓ ℓ )m T m−R ; (ℓ ′ , k ′ ) ∈ Lℓ,k .
It is easy toPsee that Vb (ℓ,k) ⊂ Ve (k,ℓ) holds. On the other hand, we observe that
any point m≥0 bm T m ∈ Ve (ℓ,k) ∩ Q[[T ]]n induces a unique parameterization
ϕ :=
(ℓ)
R−1
X
(ℓ)
m
+
σk (a(ℓ)
m )T
m=mℓ
X
bm−R T m
m≥R
of a branch of V lying above 0, and hence a vector of Puiseux series
x :=
R−1
X
(ℓ)
(ℓ)
− e1
σk (a(ℓ)
m )σk (λℓ
ℓ
m
) m E eℓ +
m=mℓ
X
(ℓ)
− e1
bm−R σk (λℓ
ℓ
m
) m E eℓ
m≥R
satisfying Fj (E, x) = 0 for j = 1, . . . , n. Then there exists (ℓ0 , k0 ) such that
P
(ℓ )
(ℓ ) (ℓ )
m/eℓ0 m/eℓ0
E
. This shows that (ℓ0 , k0 ) belongs
x = m≥mℓ σk00 (am0 )σk00 (λ−1
ℓ0 )
0
to Lℓ,k and
X (ℓ )
(ℓ)
(ℓ0 ) −1/eℓ0 m (ℓ) 1/eℓ m m 0)
ϕ = σk (λℓ )T eℓ ,
) σk (λℓ ) T
)σ
σk00 (a(ℓ
m
k0 (λℓ0
m≥mℓ
holds. Then
X
m≥R
bm T m−R =
X
(ℓ )
(ℓ )
−1/eℓ0 m (ℓ) 1/eℓ m m−R
,
) σk (λℓ ) T
0
0)
σk00 (a(ℓ
m )σk0 (λℓ0
m≥R
which shows identity (4.7).
88
(ℓ,k)
(ℓ,k)
Let us observe that G1 , . . . , Gn are obtained from F1 , . . . , Fn by ap(ℓ,k)
plying the mapping ΨR : Q[E, X] → Q[T, Y ] defined by
(ℓ,k)
ΨR
F (E, X) := T
αF
F
(ℓ)
R−1
X
m=mℓ
(ℓ)
m
R
σk (a(ℓ)
,
)T
+
Y
T
m
(ℓ,k)
where αF ∈ Z is chosen such that the order in T of ΨR (F ) is zero. In order
(ℓ,k)
to “invert” the mapping ΨR , up to a power of E, we introduce the following
(ℓ,k)
morphism ΦR : Q(T )[Y ] → Q(E)[X] of Q–algebras:
(ℓ,k)
ΦR
F (T, Y ) := F E, E
(ℓ,k)
We have E −αF ΦR
(ℓ,k)
Lemma 4.2. G1
−R
X−
R−1
X
m=mℓ
(ℓ)
m
)E
σk (a(ℓ)
m
.
(ℓ)
(ℓ,k)
ΨR (F ) = F (σk (λℓ )E eℓ , X) for any F ∈ Q[E, X].
(ℓ,k)
, . . . , Gn
form a regular sequence of Q[T, Y ].
(ℓ,k)
(ℓ,k)
Proof. Arguing by contradiction, assume that G1 , . . . , Gn do not form
(ℓ,k)
a regular sequence. Then there exists j ≥ 2 such that Gj
is a zero divisor
(ℓ,k)
(ℓ,k)
e Pe1 , . . . , Pej−1 ∈ Q[T, Y ] such
of Q[T, Y ]/(G1 , . . . , Gj−1 ), i.e. there exist H,
that
j−1
j−1
X
X
(ℓ,k)
(ℓ,k)
(ℓ,k)
(ℓ,k)
e
e
(4.8)
Pei ΨR (Fi )
Pei Gi
=
=
HΨ
R (Fj ) = HGj
i=1
i=1
e 6∈ (G1(ℓ,k) , . . . , G(ℓ,k) ). Applying the morphism Φ(ℓ,k)
holds in Q[T, Y ] with H
j−1
R
to the left and right–hand side members of identity (4.8) and multiplying by
a suitable power of E, we deduce that there exist H, P1 , . . . , Pj−1 ∈ Q[E, X]
such that
j−1
X
(ℓ)
(ℓ)
eℓ
(4.9)
Pi Fi σk (λℓ )E eℓ , X
HFj σk (λℓ )E , X =
i=1
holds in Q[E, X]. Identity (4.9) may be rewritten in the following way:
eX
ℓ −1
h=0
E
h
(ℓ)
Hh Fj σk (λℓ )E eℓ , X
=
eX
ℓ −1
h=0
E
h
j−1
X
i=1
(ℓ)
Pi,h Fi σk (λℓ )E eℓ , X ,
(4.10)
89
with Hh , P1,h , . . . , Pj−1,h ∈ Q[E eℓ , X] for 0 ≤ h ≤ eℓ − 1. Then identity (4.10)
holds if and only if the following identity
j−1
X
(ℓ)
(ℓ)
Hh Fj σk (λℓ )E eℓ , X =
Pi,h Fi σk (λℓ )E eℓ , X
i=1
holds in Q[E, X] for 0 ≤ h ≤ eℓ − 1, with at least one polynomial Hh ∈
/
(ℓ)
(ℓ)
eℓ
eℓ
(F1 (σk (λℓ )E , X), . . . , Fn (σk (λℓ )E , X)). This implies that Fj is a zero
divisor of the Q–algebra Q[E, X]/(F1 , . . . , Fj−1 ), which contradicts our hypotheses.
Let us remark that Lemma 4.2 shows in particular that the ring Q[T, Y ]/I (ℓ,k)
is Cohen–Macaulay.
Let 1 ≤ ℓ ≤ g, 1 ≤ k ≤ fℓ . From now on we fix the notations: JF :=
(ℓ,k) ∂Gi
∂Fi
,
J
.
:=
det
det ∂X
(ℓ,k)
G
∂Yj
1≤i,j≤n
j 1≤i,j≤n
Lemma 4.3. The ideal I (ℓ,k) is a radical ideal of Q[T, Y ].
Proof. Since by hypothesis the morphism π is generically unramified, the Jacobian determinant JF is not a zero divisor of Q[E, X]/(F1 , . . . , Fn ). We claim
that the Jacobian determinant JG(ℓ,k) is not a zero divisor of Q[T, Y ]/I (ℓ,k) .
e Pe1 , . . . , Pen ∈ Q[T, Y ] such that
Suppose that there exist polynomials H,
e G(ℓ,k) =
HJ
n
X
i=1
(ℓ,k)
Pei Gi
(4.11)
(ℓ,k)
(ℓ)
holds in Q[T, Y ]. Observe that JF σk (λℓ )E eℓ , X = E α ΦR (JG(ℓ,k) ) holds
for a suitable α ∈ Z. Arguing as in the proof of Lemma 4.2 we conclude that
there exist polynomials Hh , Pi,h ∈ Q[E eℓ , X] for every 0 ≤ h ≤ eℓ − 1 and
1 ≤ i ≤ n such that identity (4.11) holds if and only if the identity
(ℓ)
Hh JF σk (λℓ )E eℓ , X
=
n
X
i=1
(ℓ)
Pi Fi σk (λℓ )E eℓ , X
holds in Q[E eℓ , X] for every 0 ≤ h ≤ eℓ − 1 where at least one polynomial Hh does not belong to (F1 , . . . , Fn ). We conclude that JF is a zero
90
divisor of Q[E, X]/(F1 , . . . , Fn ), contradicting thus the hypothesis on the
generic unramifiedness of π. We conclude that JG(ℓ,k) is not a zero divisor of
Q[T, Y ]/I (ℓ,k) . This implies that the ideal generated by the n × n minors of
(ℓ,k)
(ℓ,k)
the Jacobian matrix of G1 , . . . , Gn with respect to T, Y1 , . . . , Yn has codimension at least 1 in Q[T, Y ]/I (ℓ,k) . Since Q[T, Y ]/I (ℓ,k) is a Cohen–Macaulay
ring, from [Eis95, Theorem 18.15] we conclude that I (ℓ,k) is radical.
4.2.2.
The unramifiedness of the morphism π (ℓ,k) at T =
0
In what follows, we shall use the following terminology: For a given polyb
nomial G ∈ Q[T, Y ] := Q[T, Y1 , . . . , Yn ], let us write G(T, Y ) = T α g(Y ) + G,
b ∈ Q[T, Y ] has at least order
where g is a nonzero polynomial of Q[Y ] and G
α+1 in T . The polynomial g(Y ) is called the initial form of G and is denoted
in(G).
Let us fix ℓ, k ∈ N with 1 ≤ ℓ ≤ g and 1 ≤ k ≤ eℓ . We are going
to show that the morphism π (ℓ,k) : V (ℓ,k) → A1 defined by π (ℓ,k) (t, y) := t
is unramified at every point of the fibre (π (ℓ,k) )−1 (0). For this purpose, we
are going to prove that for any point b ∈ (π (ℓ,k) )−1 (0) there exists a unique
holomorphic branch of the curve V (ℓ,k) passing through b, and b ∈ (π (ℓ,k) )−1 (0)
has multiplicity 1 in this branch. This is equivalent to showing that the
zero–dimensional affine variety defined by the (initial) ideal in(I (ℓ,k) ) ⊂ Q[Y ]
generated by the set {in(F ) : F ∈ I (ℓ,k) } has as many points as the number of
holomorphic branches of V (ℓ,k) passing through points of (π (ℓ,k) )−1 (0), namely
#(Ve (ℓ,k) ∩ Q[[T ]]n ) with the notations of Lemma 4.1. This is the content of
our next result.
Proposition 4.4. Let W (ℓ,k) denote the affine subvariety of An defined by
the ideal in(I (ℓ,k) ). Then the following identity holds in An :
(ℓ ′ )
(ℓ ′ )
(ℓ ′ )
− e1
W (ℓ,k) = {σk ′ (aR )σk ′ (λℓ ′
ℓ′
(ℓ)
1
e
)R σk (λℓ ℓ )R : (ℓ ′ , k ′ ) ∈ Lℓ,k }.
−
1
1
c (ℓ,k) := {σ (ℓ′ ′ ) (a(ℓ ′ ) )σ (ℓ′ ′ ) (λ ′ eℓ ′ )R σ (ℓ) (λ eℓ )R : (ℓ ′ , k ′ ) ∈ Lℓ,k }. We
Proof. Let W
R
ℓ
k
k
k
ℓ
c (ℓ,k) holds.
want to show that W (ℓ,k) = W
91
c (ℓ,k) . Let b ∈ W
c (ℓ,k) and let
We first prove the inclusion W (ℓ,k) ⊃ W
F ∈ I (ℓ,k) . Then there exists (ℓ′ , k ′ ) ∈ Lℓ,k such that
(ℓ ′ )
(ℓ ′ )
(ℓ ′ )
−1/eℓ ′ R (ℓ) 1/eℓ R
) σk (λℓ )
b = σk ′ (aR )σk ′ (λℓ ′
holds. Let us write F = T α in(F ) + Fb, with in(F ) ∈ Q[Y ] \ {0} and Fb ∈
Q[T, Y ] of order at least α + 1 in T . From Lemma 4.1 we have
1
− e1
P
e
(ℓ)
(ℓ ′ )
(ℓ ′ ) (ℓ ′ )
0 = F T, m≥R σk ′ (am )σk ′ (λℓ ′ ℓ ′ )m σk (λℓ ℓ )m T m−R
1
− e1
e
(ℓ)
(ℓ ′ )
(ℓ ′ ) (ℓ ′ )
= T α in(F ) σk ′ (aR )σk ′ (λℓ ′ ℓ ′ )R σk (λℓ ℓ )R + T α+1 fb(T )
= T α in(F )(b) + T α+1 fb(T ),
with fb ∈ Q[[T ]]. Then in(F )(b) = 0, which shows the inclusion W (ℓ,k) ⊃
c (ℓ,k) .
W
In order to prove the converse inclusion, let U ∈ Q[X] be a linear form for
V = V (F1 , . . . , Fn ) ⊂ An such that the minimal polynomial mu ∈ Q(E)[Z] of
the element induced by U in the extension Q[E] → Q[V ] has degree D = deg π
(for π : V → A1 (C) defined by π(ε, x) := ε; see Section 2.4).
For 1 ≤ i ≤ D, let U (i) be the element of Q(E)∗ defined by U (i) := U (x(i) ),
where {x(1) , . . . , x(D) } denote the classical Puiseux expansions of the branches
of V lying above
Q 0. Let u(i)be the rational function induced by U in Q(V∗).
Q(E) –
Observe that D
i=1 (Z − U ) annihilates u in the zero–dimensional
QD
∗
(i)
algebra Q(E) ⊗ Q(V ). Taking into account that i=1 (Z − U ) belongs
to
[Duv89]) and has degree D in Z, we conclude that mu =
QDQ(E)[Z] (see
(i)
(j)
6= U (k) for 1 ≤ j < k ≤ D and
i=1 (Z − U ) holds. This shows that U
we have the following expression for mu (Z) in Q(E)∗ [Z] (compare [Duv89]):
mu (E, Z) =
g fℓ e ℓ
Y
YY
Z−
ℓ=1 k=1 j=1
where
(ℓ)
(ℓ)
σ 1 , . . . , σ fℓ
X
m≥mℓ
(ℓ) − e1ℓ m jm m (ℓ)
)
σk (λℓ ) ξℓ E eℓ ,
U σk (a(ℓ)
m
range over all the morphisms of the Galois group of the
−1/e
and a
field extension Q ֒→ K (ℓ) and λℓ ℓ , ξℓ denote an eℓ –th root of λ−1
ℓ
primitive eℓ –th root of 1. From [Duv89, Theorem 2], we deduce that, for
1 ≤ ℓ ≤ g,
fℓ
fℓ e ℓ 1
Y
Y
Y
X
(ℓ) (ℓ) (ℓ) − eℓ m jm em
(ℓ)
(ℓ,k)
Z−
mu :=
U σk (am ) σk (λℓ ) ξℓ E ℓ
mu :=
k=1
k=1
j=1
m≥mℓ
(4.12)
92
(ℓ,k)
is an irreducible polynomial of Q((E))[Z], and, for 1 ≤ k ≤ fℓ , mu
irreducible element of Q((E))[Z] satisfying
mu(ℓ,k)
(ℓ)
σk (λℓ )T eℓ , Z
=
eℓ Y
Z−
j=1
X
U
m≥mℓ
j m
(ℓ)
σk (a(ℓ)
m ) (ξℓ T )
.
is an
(4.13)
For 1 ≤ ℓ ≤ g and 1 ≤ k ≤ fℓ , let us consider the morphism of Q-algebras
e (ℓ,k) : Q((E))[X] −→ Q((T ))[Y ]
Ψ
R
F (E, X) 7−→ F
(ℓ)
σk (λℓ )T eℓ
,
R−1
X
m=mℓ
(ℓ)
m
+ Y TR .
σk (a(ℓ)
m )T
Let us fix ℓ ′ , k ′ with 1 ≤ ℓ ′ ≤ g and 1 ≤ k ′ ≤ eℓ . Applying the morphism
e (ℓ,k) to the polynomial mu(ℓ ′ ,k ′ ) (E, U (X)), from identity (4.12) we obtain:
Ψ
R
eℓ ′ R−1
X
m
Y
(ℓ)
(ℓ,k)
(ℓ ′ ,k ′ )
e
)
T + U (Y )T R −
U σk (a(ℓ)
E, U (X) =
ΨR
mu
m
−
X
m≥mℓ
j=1
m=mℓ
1
meℓ − e1
′ ) (ℓ ′ )
eℓ m jm
(ℓ ′ )
eℓ ′
ℓ ′ m (ℓ)
T
(λ
)
ξ
)
σ
(λ
)
σ
.
U σk ′ (a(ℓ
m
ℓ
ℓ
k
k′
ℓ′
e (ℓ,k) mu(ℓ ′ ,k ′ ) E, U (X) have order
This identity shows that all the factors of Ψ
R
at most R and the coefficient of the least
nonzero power of T arising in the
(ℓ,k)
(ℓ ′ ,k ′ )
e
Laurent series ΨR mu
E, U (X) ∈ Q[Y ]((T )) is
– either of the form
− e1
− e1
(ℓ ′ )
(ℓ ′ ) (ℓ ′ )
ℓ R
ℓ ′ R (ℓ)
αU Y − σk ′ (aR )σk ′ (λℓ ′ ) σk (λℓ )
with α ∈ Q \ {0}, in case that (ℓ ′ , k ′ ) ∈ Lℓ,k holds,
– or a nonzero constant α ∈ Q.
We deduce that the coefficients of the least nonzero power of T arising in the
following elements of Q[Y ]((T )):
Y
Y
(ℓ ′ ,k ′ )
(ℓ ′ ,k ′ )
e (ℓ,k)
e (ℓ,k)
Ψ
m
,
m
E,
U
(X)
,
Ψ
E,
U
(X)
u
u
R
R
(ℓ ′ ,k ′ )∈Lℓ,k
(ℓ ′ ,k ′ )∈L
/ ℓ,k
93
Q
are of the form α b∈W
c (ℓ,k) U (Y − b) ∈ Q[Y ] with α ∈ Q \ {0}, and a constant
α
e ∈ Q \ {0} respectively. We conclude that the following identity holds:
Y
(ℓ,k)
in ΨR mu (E, U (X)) = α
U (Y − b).
(4.14)
c (ℓ,k)
b∈W
Since mu (E, U (X)) ∈ I(V ), we conclude that there exists α ∈ Z such that
(ℓ,k)
(ℓ,k)
T α ΨR mu (E, U (X)) ∈ I (ℓ,k) . Then in ΨR mu (E, U (X)) ∈ in(I (ℓ,k) ).
Now, let U1 , . . . , Un be Q–linearly independent generic linear forms. Repeating the previous arguments with U1 , . . . , Un , from identity (4.14) we conclude that W (ℓ,k) is a zero–dimensional subvariety of An . Furthermore, we
have
Y
c (ℓ,k) ≤ deg W (ℓ,k) ≤ deg
c (ℓ,k) ).
deg W
U (Y − b) = #(W
c (ℓ,k)
b∈W
c (ℓ,k) ) = #(W (ℓ,k) ). Therefore, taking into account the
This shows that #(W
c (ℓ,k) ⊂ W (ℓ,k) , we have that W (ℓ,k) = W
c (ℓ,k) holds.
inclusion W
Now we exhibit a flatness condition which assures that any point of the
(ℓ,k)
(ℓ,k)
fibre (π (ℓ,k) )−1 (0) is (G1 , . . . , Gn )–smooth. For this purpose, we introduce the notion of standard basis (see [CLO98]). In our setting, a set
{G1 , . . . , Gs } ⊂ Q[T,Y ] := Q[T,Y1 , . . . ,Yn ] is called a standard basis (of the
ideal I they generate) if the ideal (in(G1 ), . . . , in(Gs )) generated by the initial
forms of G1 , . . . , Gs in Q[Y ] agrees with the ideal in(I) := (in(G) : G ∈ I)
generated by the initial forms of all the polynomials G ∈ I.
Theorem 4.5. Let notations and assumptions be as above. Suppose further
(ℓ,k)
(ℓ,k)
that G1 (T, Y ), . . . , Gn (T, Y ) form a standard basis of the ideal I (ℓ,k) .
Then the Jacobian determinant JG(ℓ,k) does not vanish at any point of
(π (ℓ,k) )−1 (0).
(ℓ,k)
(ℓ,k)
form a standard basis of I (ℓ,k) we see that
(ℓ,k)
(ℓ,k)
(π (ℓ,k) )−1 (0) = {0} × V G1 (0, Y ), . . . , Gn (0, Y ) =
(ℓ,k)
(ℓ,k) = {0} × V in(G1 ), . . . , in(Gn ) = {0} × W (ℓ,k)
Proof. Since G1
, . . . , Gn
holds. From Proposition 4.4 we see that for any b ∈ W (ℓ,k) there exists a
(ℓ,k)
unique vector of power series ϕ ∈ Q[[T ]] such that ϕ(0) = b and Gi (T, ϕ) =
94
0 hold for 1 ≤ i ≤ n. Then [ANMR91, Lemma 3] shows that Y = b has
(ℓ,k)
(ℓ,k)
multiplicity 1 as a zero of the ideal generated by G1 (0, Y ), . . . , Gn (0, Y ).
Therefore, JG(ℓ,k) does not vanish at any point (0, y) ∈ (π (ℓ,k) )−1 (0).
4.2.3.
A global Newton–Hensel lifting
In the context of this chapter, the projection problem (see Section 2.4)
can be stated as follows:
e (1) ), . . . , (Ee(g) , X
e (g) )} of paLifting of a projection: given a set {(Ee(1) , X
rameterizations of V (i.e., by their singular parts), whose orbits under the
action of the Galois group of Q/Q form a system of rational Puiseux expansions of the branches of the curve V lying above 0, and a generic linear form
U ∈ Q[X], compute the projection polynomial mu ∈ Q(E)[Z].
Let us fix ℓ with 1 ≤ ℓ ≤ g and let S1 , S2 be indeterminates over Q. Let
q be a monic irreducible polynomial of Q[S1 ] of degree fℓ = [K (ℓ) : Q] such
that there exists a Q–isomorphism of fields Υℓ : Q[S1 ]/(q (ℓ) (S1 )) → K (ℓ) .
(ℓ)
For any m ≥ mℓ and 1 ≤ j ≤ n, let f (ℓ) , fm,j be the (unique) polynomials of
(ℓ)
(ℓ)
Q[S1 ] of degree at most fℓ −1, such that Υℓ (f (ℓ) ) := λ−1
ℓ and Υℓ (fm ) := am,j .
Finally, let p(ℓ) ∈ Q[S1 , S2 ] be the polynomial p(ℓ) := S2eℓ − f (ℓ) (S1 ), and let
(ℓ)
W (ℓ) := {(s1 , s2 ) ∈ A2 : p(ℓ) (s1 , s2 ) = 0, q (ℓ) (s1 ) = 0}.
(4.15)
It is easy to see that W (ℓ) is a zero–dimensional variety of degree deg W (ℓ) =
eℓ fℓ . [Duv89] shows that the field K (ℓ) is the field extension of Q generated
(ℓ)
by the coefficients aj,m for 1 ≤ j ≤ n and mℓ ≤ m < R. In particular, K (ℓ)
is the minimal field extension of Q containing the coefficients of the singular
parts of the given set of rational Puiseux expansions.
Pκ
(ℓ)
m
For κ ≥ R, let u(κ,ℓ) :=
∈ Q(S1 , S2 , T ) and
m=mℓ U (fm (S1 ))(S2 T )
let χu(κ,ℓ) ∈ Q(T )[Z] denote the characteristic polynomial of the projection
(ℓ)
(ℓ)
πu(κ,ℓ) : A1 × W (ℓ) → A2 defined by πu(κ,ℓ) (t, s1 , s2 ) := (t, u(κ,ℓ) (t, s1 , s2 )). We
have
χu(κ,ℓ) =
fℓ e ℓ Y
Y
k=1 j=1
Z−
κ
X
m=mℓ
j (ℓ) − e1ℓ m (ℓ)
)
ξℓ σk (λℓ )T
U σk (a(ℓ)
.
m
(ℓ)
(4.16)
−1/eℓ
Observe that if the norm in the field extension Q(T eℓ ) → K (ℓ) (σk (λℓ
)T )
95
is extended to polynomials, then χu(κ,ℓ) is the norm of Z−
This shows that χu(κ,ℓ) is an element of Q(T eℓ )[Z].
Pκ
(ℓ) −1/eℓ m
T .
m=mℓ U (am )λℓ
e ∈
Before continuing, we introduce the following terminology: for G, G
e
Q((E)) and any s ∈ Z, we say that G approximates G with precision s in
e has order at least s + 1 in E. We shall use
Q((E)) if the Laurent series G − G
e mod (E s+1 ). Furthermore, if G, G
e are two elements of a
the notation G ≡ G
e approximates G with precision s if
polynomial ring Q((E))[Z], we say that G
e approximates the corresponding coefficient
every coefficient e
a ∈ Q((E)) of G
a ∈ Q((E)) of G with precision s (in the sense of the previous definition).
From identities (4.12) and (4.16) we easily deduce that the congruence
relation
eℓ
κ−δ0 mℓ eℓ fℓ +1
m(ℓ)
)
(4.17)
u (T , Z) ≡ χu(κ,ℓ) (T, Z) mod (T
holds in Q((T ))[Z], with δ0 := −1 for mℓ < 0 and δ0 := 0 otherwise. Taking
into account that χu(κ,ℓ) (T, Z) is an element of Q(T eℓ )[Z], replacing T eℓ by E
in (4.17) we obtain the following result.
Lemma 4.6. For any κ ≥ R, χu(κ,ℓ) (E 1/eℓ , Z) ∈ Q(E)[Z] approximates the
(ℓ)
polynomial mu (E, Z) ∈ Q((E))[Z] with precision ⌊ κ−δ0emℓ ℓ eℓ fℓ ⌋ in Q((E))[Z].
Now we state our version of the global Newton–Hensel lifting.
Theorem 4.7. Let the assumptions of Theorem 4.5 hold. Let be given κ ≥ 0.
(ℓ)
For 1 ≤ j ≤ n, let Gj be the following element of Q[S1 , S2 , S2−1 , T, Y ]:
(ℓ)
Gj (S1 , S2 , T, Y
) := T
αjℓ
eℓ
Fj T ,
R−1
X
(ℓ)
fm
(S1 )(S2 T )m
+YT
m=mℓ
(ℓ)
R
.
(ℓ)
Let NG(ℓ) be the Newton–Hensel operator associated to G1 , . . . , Gn , namely
 (ℓ) 
 
G1
Y1
(ℓ)
∂G −1
 .. 
 .. 
i
· . 
NG(ℓ) (Y ) :=  .  −
(4.18)
∂Yj 1≤i,j≤n
(ℓ)
Yn
Gn
and let NGκ (ℓ) denote the κ–th fold iteration of NG(ℓ) . Finally, let
u
e(κ,ℓ) := U
R−1
X
m=mℓ
(ℓ)
(ℓ)
fm
(S1 )(S2 T )m + NGκ (ℓ) S1 , S2 , T, fR (S1 )S2R T R
96
1
and let χue(ℓ,k) ∈ Q(T )[Z] be its characteristic polynomial. Then χue(ℓ,k) (E eℓ ,Z)
κ
(ℓ)
ℓ eℓ fℓ +2
approximates the polynomial mu with precision ⌊ R−1−δ0 m
⌋ in Q((E))[Z].
eℓ
Proof. Let (s1 , s2 ) be a point of the variety W (ℓ) . Then there exists a (unique)
(ℓ)
pair (k, j) with 1 ≤ k ≤ fℓ and 1 ≤ j ≤ eℓ , such that f (ℓ) (s1 ) = σk (λ−1
ℓ ),
(ℓ)
(ℓ) (ℓ)
j (ℓ) −1/eℓ
) and fm (s1 ) = σk (am ) hold for mℓ ≤ m ≤ R. This
s2 = ξℓ σk (λℓ
implies that the following identity holds in Q[T, Y ] for 1 ≤ i ≤ n:
α
(ℓ)
(ℓ,k)
Gi (s1 , s2 , T, Y ) = s2 jℓ Gi
(ℓ)
(s2 T, s−R
2 Y ).
(4.19)
(ℓ)
n
Let us observe that sR
2 σk (aR ) ∈ A belongs to the affine variety defined by
(ℓ)
(ℓ)
G1 (s1 , s2 , 0, Y ), . . . , Gn (s1 , s2 , 0, Y ). Furthermore, from Theorem 4.5 and
(ℓ)
identity (4.19) we conclude that JG(ℓ) (T, Y ) := det(∂Gi /∂Yj )1≤i,j≤n (s1 , s2 , T, Y )
(ℓ) (ℓ)
(ℓ) (ℓ)
n+1
does not vanish at (0, sR
, and hence JG(ℓ) (T, sR
2 σk (aR )) is
2 σk (aR )) ∈ A
a unit in the local ring (Q[[T ]], (T )).
From Hensel’s Lemma (see e.g. [Eis95]) in the version of [HKP+ 00] we
deduce that the following congruence relation holds in Q[[T ]]n :
X (ℓ)
κ
(ℓ) (ℓ)
(ℓ) m m−R
)s2 T
mod (T 2 ).
≡
σk (am
NGκ (ℓ) s1 , s2 , T, σk (aR )sR
2
m≥R
Therefore, we obtain
R−1
X
R
(ℓ) (ℓ)
(ℓ)
≡
U
fm
(s1 )(s2 T )m + NG(ℓ) s1 , s2 , T, σk (aR )sR
2 T
m=mℓ
X
κ
(ℓ) (ℓ)
m
≡ U
mod (T R+2 ),
σk (am )(s2 T )
m≥mℓ
κ
κ
which implies u
e(κ,ℓ) (s1 , s2 , T ) ≡ u(R−1+2 , ℓ) (s1 , s2 , T ) mod (T R+2 ).
(ℓ)
Lemma 4.6 shows that χu(R−1+2κ, ℓ) (E 1/eℓ , Z) approximates mu in Q((E))[Z]
κ
ℓ eℓ fℓ +2
with precision ⌊ R−1−δ0 m
⌋. Therefore, χue(κ, ℓ) (E 1/eℓ , Z) also approxieℓ
(ℓ)
κ
ℓ eℓ fℓ +2
⌋. This proves the
mates mu in Q((E))[Z] with precision ⌊ R−1−δ0 m
eℓ
theorem.
4.2.4.
Unramifiedness and flatness conditions
All the hypotheses of Theorems 4.5 and 4.7 are fairly “geometric” in nature, and hence reasonable assumptions from our point of view (compare
97
[HKP+ 00]), except perhaps for the standard basis requirement. Nevertheless,
this is not an arbitrary “algebraic” requirement, as shown by the following
result.
Lemma 4.8. Let notations and assumptions be as in Lemmas 4.1, 4.2 and
(ℓ,k)
4.3. Suppose that the morphism π (ℓ,k) is unramified at T = 0. Then G1 , . . . ,
(ℓ,k)
Gn form a standard basis of the ideal I (ℓ,k) .
Proof. Let x ∈ An+1 be a point of the fibre (π (ℓ,k) )−1 (0). Let (OV (ℓ,k) , x , mx )
denote the local ring of the point x on the variety V (ℓ,k) and let (OA1 , 0 , m0 )
denote the local ring of 0 on A1 . Since π (ℓ,k) is unramified at T = 0, we have
mx = (π (ℓ,k) )∗ (m0 )
(4.20)
for any x ∈ (π (ℓ,k) )−1 (0), where (π (ℓ,k) )∗ denotes the local homomorphism
(π (ℓ,k) )∗ : OA1 , 0 → OV (ℓ,k) , x induced by the morphism π (ℓ,k) .
Identity (4.20) implies that the morphism dx π (ℓ,k) : TV (ℓ,k) , x → TA1 , 0 of
tangent spaces is injective [Dan94]. We deduce that the dimension dim(TV (ℓ,k) , x )
of the tangent space TV (ℓ,k) , x of V (ℓ,k) at x is at most 1. Taking into account
that V (ℓ,k) is an equidimensional variety of dimension 1 (Lemma 4.2), we
conclude that dim(TV (ℓ,k) , x ) = 1. Therefore, x is a smooth point of V (ℓ,k) .
Identity (4.20) shows that the quotient ring OV (ℓ,k) , x /(π (ℓ,k) )∗ (m0 ) is a
zero–dimensional Q–algebra. Let us observe that OV (ℓ,k) , x is a Cohen–Macaulay
ring (because it is a localization of a Cohen–Macaulay ring), the local ring
OA1 , 0 is a regular ring and the identity
dim OV (ℓ,k) , x = dim OA1 , 0 + dim OV (ℓ,k) , x /(π (ℓ,k) )∗ (m0 )
holds. Then applying [Mat86, Theorem 23.1] we conclude that the local
homomorphism
(π (ℓ,k) )∗ : OA1 , 0 → OV (ℓ,k) , x
(4.21)
induced by π (ℓ,k) is flat.
We observe that the localization Q[V (ℓ,k) ]m0 is a semilocal ring, whose
maximal ideals correspond to the maximal ideals mx induced by the points x
of (π (ℓ,k) )−1 (0). Therefore, since the morphism of (4.21) is flat for any point
x ∈ of (π (ℓ,k) )−1 (0), applying [Mat86, Theorem 7.1] we conclude that
(π (ℓ,k) )∗ : Q[A1 ]m0 → Q[V (ℓ,k) ]m0
98
induced by π (ℓ,k) is flat, i.e. π (ℓ,k) is flat at T = 0. Therefore, from [Art76,
Part I, Proposition 3.1] (see also [BM93]) it follows that any syzygy (h1 , . . . ,
(ℓ,k)
(ℓ,k)
hn ) ∈ Q[Y ]n of the polynomials G1 (0, Y ), . . . , Gn (0, Y ) “lifts” to a syzygy
(ℓ,k)
(ℓ,k)
(H1 , . . . , Hn ) ∈ Q[T, Y ]n of G1 , . . . , Gn , i.e. for 1 ≤ i ≤ n the identity
Hi (0, Y ) = hi (Y ) holds.
Now we adapt the contents of e.g. [MPT92] to our setting. For F ∈
Q[T, Y ], let oT (F ) denote the highest power of T dividing F . We claim that
any polynomial G ∈ I (ℓ,k) has a representation
G=
n
X
(ℓ,k)
Hi Gi
(4.22)
i=1
with order oT (Hi ) ≥ oT (G) for 1 ≤ i ≤ n. Let G ∈ I (ℓ,k) be a polynomial
P
(ℓ,k)
with a representation G = ni=1 Hi Gi . Let α := min{oT (Hi ) : 1 ≤ i ≤ n},
and suppose that α < oT (G) holds. Let J be the set of indices i for which
α = oT (Hi ) holds. Then the identity
X
(ℓ,k)
(T −α Hi )(0, Y )Gi (0, Y ) = 0
i∈J
shows that (h1 , . . . , hn ) ∈ Q[Y ]n , with hi := (T −α Hi )(0, Y ) if i ∈ J and
(ℓ,k)
(ℓ,k)
hi := 0 otherwise, is a syzygy of G1 (0, Y ), . . . , Gn (0, Y ). Then there
e1, . . . , H
e n ) ∈ Q[T, Y ]n of the syzygy (h1 , . . . , hn ), and we
exists a lifting (H
have:
n
X
e i )G(ℓ,k) ,
G=
(Hi − T α H
i
i=1
e i ) > α for 1 ≤ i ≤ n. Repeating this argument at most
with oT (Hi − T α H
oT (G) times, we conclude the validity of our claim.
Finally, let G ∈ I (ℓ,k) . Then we have a representation of G as in (4.22),
with order oT (Hi ) ≥ oT (G) for 1 ≤ i ≤ n. Let J be the (nonempty) set of
indices i for which oT (G) = oT (Hi ) holds. Then we have
X
(ℓ,k)
T −oT (G) Hi (0, Y ) · Gi (0, Y )
in(G) = T −oT (G) G (0, Y ) =
i∈J
X
(ℓ,k)
=
T −oT (G) Hi (0, Y ) · in(Gi ).
i∈J
(ℓ,k)
This shows that G1
(ℓ,k)
, . . . , Gn
form a standard basis of the ideal I (ℓ,k) .
99
4.3. ALGORITHMS AND COMPLEXITY ESTIMATES
4.3.
Algorithms and complexity estimates
Let notations and assumptions be as in Section 4.1. Let δ := deg V
denote the degree of the variety V , and let D := deg π denote the degree
of the morphism π : V → A1 . Suppose that we are given a straight–line
program β computing F1 , . . . , Fn with T arithmetic operations in Q.
Let S1 , S2 be indeterminates over Q. With the notations of Section 4.2.3,
(ℓ)
(ℓ)
for 1 ≤ ℓ ≤ g and mℓ ≤ m ≤ R, let q (ℓ) , f (ℓ) , fm,1 , . . . , fm,n ∈ Q[S1 ] and p(ℓ) ∈
Q[S1 , S2 ] be polynomials defining the system of rational Puiseux expansions
of the branches of V lying above 0 of Section 4.2.3. In particular, we have
(ℓ)
the estimates deg(q (ℓ) ) = fℓ , deg(f (ℓ) ) < fℓ and deg(fm, i ) < fℓ for 1 ≤ i ≤ n,
and the singular parts of the (classical) Puiseux expansions of the branches
of V lying over 0 are given by
g n
[
ℓ=1
(ℓ)
eℓ
T ,
R
X
m=mℓ
(ℓ)
o
(ℓ)
(ℓ)
(ℓ)
m
;
p
(s
,
s
)
=
q
(s
)
=
0
,
fm
(s1 )sm
T
1
2
1
2
(4.23)
(ℓ)
where fm := (fm,1 , . . . , fm,n ) ∈ Q[S1 ]n . Let U ∈ Q[X] a generic linear
form, i.e. a linear form whose projection polynomial mu ∈ Q(E)[Z] satisfies
degZ mu = D. Then identity (4.12) of Section 4.2 shows that mu has the
following factorization into irreducible factors in Q((E))[Z]:
mu =
g
Y
ℓ=1
m(ℓ)
u :=
g fℓ e ℓ Y
YY
ℓ=1
k=1 j=1
1
X
(ℓ) (ℓ) (ℓ) − eℓ m jm em
.
Z − U σk (am ) σk (λℓ ) ξℓ E ℓ
m≥mℓ
(4.24)
In this section we exhibit an algorithm which has as input the straight–
(ℓ)
(ℓ)
line program β and the dense representation of p(ℓ) , q (ℓ) , f (ℓ) , fm,1 , . . . , fm,n
for 1 ≤ ℓ ≤ g and mℓ ≤ m ≤ R and computes a geometric solution of V .
Let us fix ℓ with 1 ≤ ℓ ≤ g. The critical part of our algorithm is a
(ℓ)
procedure which computes a suitable approximation m̂u ∈ Q(E)[Z] of the
(ℓ)
polynomial mu ∈ Q((E))[Z]. This procedure applies our variant of the global
Newton–Hensel lifting based on Theorem 4.7. For this purpose, we shall deal
with the variety W (ℓ) of (4.15), namely
W (ℓ) := {(s1 , s2 ) ∈ A2 : q (ℓ) (s1 ) = 0, p(ℓ) (s1 , s2 ) = 0}.
100
From the fact that deg W (ℓ) = eℓ fℓ holds, we easily conclude that S2 is a
primitive element of the Q–algebra extension Q ֒→ Q[W (ℓ) ]. Therefore, we
have a geometric solution of W (ℓ) of the form
(ℓ)
(ℓ)
W (ℓ) = {(s1 , s2 ) ∈ A2 : mS2 (s2 ) = 0, s1
∂mS2
(s2 ) − v (ℓ) (s2 ) = 0},
∂Z
(4.25)
(ℓ)
where mS2 ∈ Q[Z] is the minimal polynomial of S2 in the extension Q ֒→
Q[W (ℓ) ] and v (ℓ) ∈ Q[Z] satisfies deg v (ℓ) < deg W (ℓ) .
Lemma 4.9. There exists an algorithm that computes the geometric solution
(4.25) of W (ℓ) in O(max{eℓ , fℓ }M(eℓ fℓ )) arithmetic operations in Q.
Proof. Let us suppose first fℓ = 1. Then we may assume without loss of
generality q (ℓ) = S1 . Furthermore, we have f (ℓ) ∈ Q\{0} and p(ℓ) = S2eℓ −f (ℓ) .
(ℓ)
Therefore, mS2 = p(ℓ) = Z eℓ − f (ℓ) and v (ℓ) = 0 yield in fact the geometric
solution of W (ℓ) we are looking for (and we have nothing to compute).
Now suppose that fℓ > 1 holds. Let us introduce a new indeterminate
Λ, and let us consider the linear form L := ΛS1 + S2 ∈ Q[Λ][S1 , S2 ]. It
is easy to see that L is a primitive element of the integral ring extension
Q[Λ] ֒→ Q[Λ] ⊗ Q[W (ℓ) ], with minimal equation
(ℓ)
mL (Z) = ResS1 q (ℓ) (S1 ), p(ℓ) (S1 , Z − ΛS1 ) ,
(4.26)
where ResS1 (f, g) denotes the resultant of f and g with respect to S1 . Arguing as in (2.10), we have a congruence relation:
!
(ℓ)
∂m
(ℓ)
(ℓ)
S2
(Z) + ve(ℓ) (Z)
mod (Λ2 ),
mL (Z) = mS2 (Z) + Λ S1
∂Z
(ℓ)
with ve(ℓ) ∈ Q[Z], deg ve(ℓ) < eℓ fℓ and S1 (∂mS2 /∂Z)(S2 ) + ve(ℓ) (S2 ) ∈ I(W (ℓ) ).
(ℓ)
v (ℓ) can be obtained from the resultant of the right–
Then mS2 and v (ℓ) := −e
hand side of identity (4.26) modulo Λ2 . Using interpolation in the variable Z,
this computation can be performed with O(max{eℓ , fℓ }M(eℓ fℓ )) arithmetic
operations in Q.
Our variant of the global Newton–Hensel lifting requires the R–th “ini(ℓ)
tial approximation” of mu given by the following expression (compare with
101
(4.24)):
eℓ
m
e (ℓ)
u (T , Z)
:=
fℓ e ℓ Y
Y
k=1 j=1
Z−
R
X
m=mℓ
U
(ℓ) − e1ℓ m jm m
(ℓ)
)
σk (λℓ ) ξℓ T
σk (a(ℓ)
m
. (4.27)
Lemma 4.10. There exists a computation tree which takes as input the poly(ℓ)
(ℓ)
nomials p(ℓ) , q (ℓ) , f (ℓ) , fm, i (1 ≤ i ≤ n, mℓ ≤ m ≤ R), mS2 , v (ℓ) , which define
the ℓ–th expansion of the given system of rational Puiseux expansions of V
and form the geometric solution (4.25) of W (ℓ) , and computes the dense rep(ℓ)
resentation of m
e u in O(Rℓ eℓ fℓ M(eℓ fℓ )) arithmetic operations in Q, where
Rℓ := (R − mℓ )eℓ fℓ + 1.
(ℓ)
Proof. From the definition of m
e u and the variety W (ℓ) we easily see that
(ℓ)
T −mℓ eℓ fℓ m
e u (T eℓ , T mℓ Z) equals the characteristic polynomial χue of the polyPR
(ℓ)
m m−mℓ
nomial u
e(T, S1 , S2 ) :=
in the Q–algebra
m=mℓ U ( fm (S1 ) ) S2 T
(ℓ) ∼
1
(ℓ)
Q[T ] ⊗ Q[W ] = Q[A × W ]. Let us observe that S2 is a primitive element
(ℓ)
of the extension Q[T ] ֒→ Q[A1 × W (ℓ) ] and the input polynomials mS2 , v (ℓ)
also yield a geometric solution of the variety A1 × W (ℓ) .
In order to compute the dense representation of χue we use
forward adaptation of the algorithm of [HMW01, Lemma 3].
(ℓ)
Q(eℓ fℓ )×(eℓ fℓ ) be the companion matrix of the polynomial mS2 .
characteristic polynomial of the matrix N := u
e(T, v (ℓ) (M ), M )
characteristic polynomial χue .
a straightLet M ∈
Then the
equals the
Let us suppose first that R = mℓ holds. Then χue is a pseudo–homogeneous
polynomial whose coefficients can be computed using [HMW01, Lemma 3]
with O(eℓ fℓ M(eℓ fℓ )) arithmetic operations in Q. On the other hand, if
R 6= mℓ , taking into account that the algorithm manipulates polynomials in
(ℓ)
T of degree at most (R − mℓ )eℓ fℓ , and the fact that the polynomial mS2 (Z)
does not depend on the variable T , we conclude that the procedure underlying [HMW01, Lemma 3] can be executed with O((R − mℓ )e2ℓ fℓ2 M(eℓ fℓ ))
arithmetic operations in Q. In conclusion, we see that the procedure takes in
both cases O(((R − mℓ )eℓ fℓ + 1)M(eℓ fℓ )) arithmetic operations in Q. Finally,
(ℓ)
taking into account that the dense representation of m
e u can be immediately
obtained from that of χue finishes the proof of the lemma.
Now we can describe the algorithm computing an arbitrary approximation
(ℓ)
in Q(E)[Z] of the polynomial mu ∈ Q((E))[Z]. This algorithm applies our
102
variant of the global Newton–Hensel lifting (Theorem 4.7), combined with
an adaptation of the procedure of [GLS01, Proposition 7]. For this purpose,
following Theorem 4.7, let Y1 , . . . , Yn be indeterminates over Q, let Y :=
(ℓ)
(ℓ)
(Y1 , . . . , Yn ), and let us define G1 , . . . , Gn ∈ Q[S1 , S2−1 , S2 , T, Y ] by:
!
R−1
X
(ℓ)
(ℓ)
Gj := T αjℓ Fj T eℓ ,
(4.28)
fm
(S1 )(S2 T )m + Y T R .
m=mℓ
Proposition 4.11. Let us fix κ > 0. Then there exists a computation tree
(ℓ)
which takes as input the polynomials p(ℓ) , q (ℓ) , f (ℓ) , fm, i (1 ≤ i ≤ n, mℓ ≤ m ≤
(ℓ)
R), mS2 , v (ℓ) , which define the ℓ–th parameterization of the given system of
rational Puiseux expansions of V and form the geometric solution (4.25) of
(ℓ)
(ℓ)
W (ℓ) , and computes an approximation m̂u ∈ Q(E)[Z] of mu in Q((E))[Z]
⌉ + 1 and parameterizations of Y1 , . . . , Yn in terms of the
with precision ⌈ R+κ
eℓ
linear form U up to order ⌈ R+κ
⌉ + 1, in
eℓ
O(n(Tℓ + n4 )M(κ + δ0 mℓ eℓ fℓ + (Rℓ − 1)eℓ fℓ )M(eℓ fℓ )
arithmetic operations in Q, where Rℓ := (R − mℓ )eℓ fℓ + 1, δ0 := −1 for
mℓ < 0 and δ0 := 0 otherwise, and Tℓ denotes arithmetic operations in Q
(ℓ)
(ℓ)
required for the evaluation of the polynomials G1 , . . . , Gn .
Proof. Theorem 4.5 shows that the Newton operator NG(ℓ) of (4.18) is well
(ℓ)
(ℓ)
. Then Theorem 4.7 shows that
defined at fR (s1 )sR
2 for any (s1 , s2 ) ∈ W
from the τ := ⌈log2 (κ + δ0 mℓ eℓ fℓ + 1)⌉–fold iteration of the Newton operator
(ℓ)
NG(ℓ) we obtain a rational function m̂u ∈ Q(E)[Z] which approximates mu
⌋.
in Q((E)) with precision ⌊ R+κ
eℓ
In order to compute m̂u (T eℓ , Z) we use an adaptation of the procedure of
[GLS01, Proposition 7]: we start with the initial approximation provided by
(ℓ)
the polynomial m
e u of (4.27) and parameterizations of X1 , . . . , Xn in terms
(ℓ)
(ℓ)
of the linear form U up to order R + 1, i.e. elements ve1 , . . . , ven of Q(E)[Z]
(ℓ)
(ℓ)
(ℓ)
eu
(T eℓ , U )Xi ≡ vei (T eℓ , U ) mod (T R+1 , mu (T eℓ , U )). Then we
such that ∂ m
∂Z
perform τ steps of the global Newton–Hensel lifting of [GLS01, Proposition
(ℓ)
(ℓ)
7] applied to the polynomials G1 , . . . , Gn .
(ℓ)
Applying Lemma 4.10 we can compute the polynomial m
e u of (4.27)
in O(Rℓ eℓ fℓ M(eℓ fℓ )) arithmetic operations in Q. Combining Lemma 4.10
103
and the formulae of e.g. [ABRW96], [Rou97] or [GLS01] as in the proof of
Lemma 4.9 we obtain the parameterizations of X1 , . . . , Xn in terms of U up
to order R + 1 using O(nRℓ eℓ fℓ M(eℓ fℓ )) arithmetic operations in Q.
(ℓ)
Now, applying [GLS01, Lemma 2] we obtain an approximation of mu
with precision R + κ + 1 in Q((T ))[Z] and parameterizations of X1 , . . . , Xn in
terms of U up to order R + κ + 1 with time
O(n(Tℓ + n4 )M(κ + δ0 mℓ eℓ fℓ + (Rℓ − 1)eℓ fℓ )M(eℓ fℓ ) .
(ℓ)
Since mu (T eℓ , Z) and the parameterizations of X1 , . . . , Xn in terms of U are
(ℓ)
elements of Q((T eℓ ))[Z], replacing T eℓ by E we obtain m̂u and the parame⌋ + 1. Adding the
terizations of X1 , . . . , Xn in terms of U up to order ⌊ R+κ
eℓ
complexity of each step of our procedure the proposition follows.
Now we state the main result of this section.
Theorem 4.12. There exists an algorithm in Q[E, X] which takes as input
the straight–line program β defining the polynomials F1 , . . . , Fn and the given
system of rational Puiseux expansions and computes a geometric solution of
V in
!
g
X
neℓ (Tℓ + n4 )(δ + δ0 mℓ fℓ + Rℓ )M(eℓ fℓ )
O
ℓ=1
arithmetic operations in Q, where Rℓ := (R − mℓ )eℓ fℓ + 1, δ0 := −1 for
mℓ < 0 and δ0 := 0 otherwise, and Tℓ denotes the number of arithmetic
(ℓ)
(ℓ)
operations in Q required for the evaluation of the polynomials G1 , . . . , Gn
of (4.28). Furthermore, for any ρ ≥ 2, such a computation tree can be
1
≥ 43 .
randomly constructed with a probability of success of at least 1 − 2ρ
Proof. Let U ∈ Q[X] be a generic linear form. Let us fix ρ ≥ 2. Using the
Zippel–Schwartz Theorem (Theorem 2.1), we conclude that the coefficients
of U can be randomly chosen in the set {1, . . . , 4ρnD2 } with a probability of
1
success of at least 1 − 2ρ
≥ 43 , where D := deg π.
Let δ := deg V . Applying Proposition 4.11 for 1 ≤ ℓ ≤ g with κ :=
(ℓ) (ℓ)
(ℓ)
3eℓ δ − R, we obtain elements m̂u , v̂1 , . . . , v̂n (1 ≤ ℓ ≤ g) of Q(E)[Z] such
that:
(ℓ)
(ℓ)
1. m̂u (E, Z) ≡ mu (E, Z) modulo (E 3δ+1 ),
104
2.
(ℓ)
∂ m̂u
∂Z
(ℓ)
(ℓ)
(E, U )Xi ≡ v̂i (E, U ) modulo E 3δ+1 , mu (E, U ) ,
(ℓ)
(ℓ)
3. degZ m̂u ≤ eℓ fℓ and degZ vi ≤ eℓ fℓ − 1 for 1 ≤ i ≤ n.
These polynomials can be computed with
O
g
X
n(Tℓ + n4 )M(eℓ δ + δ0 mℓ eℓ fℓ + (Rℓ − 1)eℓ fℓ )M(eℓ fℓ )
ℓ=1
!
arithmetic operations in Q. Let v1 , . . . , vn be the elements of Q(E)[Z] parameterizing X1 , . . . , Xn in terms of the linear form U in V , i.e. satisfying
∂mu
(E, U )Xi ≡ vi (E, U ) mod I(V ) for 1 ≤ i ≤ n. From [Sch03, Proposition
∂Z
1] we see that the orders oE (mu ), oE (v1 ), . . . , oE (vn ) are bounded from below
by −δ. Combining this observation with properties (1), (2), (3) we conclude
that the following congruence relations hold in Q((E))[Z]:
Q
(ℓ)
m̆u := gℓ=1m̂u ≡ mu mod (E 2δ+1 ),
P
Q
(ℓ)
(ℓ′ )
v̂i
≡ vi mod (E 2δ+1 ).
v̆i := 1≤ℓ≤g
ℓ′ 6=ℓ m̂u
Using fast procedures for multiplication and Chinese Remainder Theorem
(see e.g. [BP94]), we compute the polynomials m̆u , v̆1 , . . . , v̆n using O(nM(δD))
arithmetic operations in Q.
Taking into account the estimates
degZ mu = D, degZ vi ≤ D − 1, (1 ≤ i ≤ n),
degE mu ≤ δ, degE vi ≤ δ (1 ≤ i ≤ n),
(see [Sch03]), we conclude that mu , v1 , . . . , vn can be computed from the truncated Laurent series m̆u , v̆1 , . . . , v̆n using Padé approximants. More precisely,
by interpolation in the variable Z we reduce the computation of the polynomials mu , v1 , . . . , vn to at most (n + 1)D problems of Padé approximation of
degree at most δ. Thus, using a fast algorithm for computing Padé approximations (see e.g. [BP94]), we conclude that the polynomials mu , v1 , . . . , vn
can be computed in O(nM(δD)) arithmetic operations in Q. Adding the
number of arithmetic operations used in each step of our procedure we deduce the complexity estimate of the statement of Theorem 4.12.
Let us make here a few remarks concerning the hypotheses and complexity estimates of Theorem 4.12. First we observe that the parameter Tℓ can be
105
roughly estimated by O(T +nRℓ ), where T is the number of arithmetic operations in Q that the straight–line program requires for computing F1 , . . . , Fn .
Then we have the rough worst–case estimate of O(n4 T δD4 ) arithmetic operations in Q for the procedure underlying Theorem 4.12. Nevertheless, these
estimates can be improved in several important cases, such as that where
R = mℓ and Q[E](E) ֒→ Q[V ](E) is an integral extension: In this case we
have the rough estimate O((T + n4 )nδDe) arithmetic operations in Q, with
e := max{eℓ : 1 ≤ ℓ ≤ g} (see Subsections 4.4.3 and 4.4.4).
Theorem 4.12 generalizes the results of [HKP+ 00] and [Sch03] in the unidimensional case. More precisely, in case that the “known” π–fibre is unramified, the corresponding (rough) estimate is of O((n4 + T )Dδ) arithmetic
operations in Q, which improve the estimates of [HKP+ 00] and have the same
asymptotic behaviour as those of [Sch03].
The algorithm underlying Theorem 4.12 proceeds by computing a suit(ℓ)
able approximation of the factors mu of the minimal polynomial mu =
Q
(ℓ)
1≤ℓ≤g mu of the linear form U . Observe that for 1 ≤ ℓ ≤ g the polynomial
(ℓ)
mu is an irreducible polynomial of Q((E))[Z] (see Section 4.2). In this sense,
this algorithm constitutes an improvement of the refinements described in
Section 3 of [HKP+ 00] (based on the factorization of the polynomial mu in
Q[E, Z]).
The singular parts (4.23) can be efficiently computed from the input polynomials F1 , . . . , Fn and a geometric solution of an unramified fibre of the
morphism π, by a suitable combination of the following algorithmic tools.
A Newton polygon algorithm for computing the singular parts of a
system of rational Puiseux expansions as in [Duv89] or [Wal00].
A projection procedure for unramified fibres as in [Sch03].
The asymptotic time complexity of such a procedure is roughly O(During +
̺2 ) arithmetic operations in Q, where ̺ denotes the geometric degree of the
system F1 , . . . , Fn (in the sense of [GHH+ 97]). Observe that the estimates
D ≤ δ ≤ ̺ hold.
In [PR11] and [PR12] the authors describe an algorithm to compute the
singular parts of rational Puiseux expansions of an equation given by a bivariate polynomial F (X, Y ) = 0. The approach is based on a modification of an
106
algorithm due to D. Duval ([Duv89]). The randomized algorithm presented
in [PR12] chooses a suitable finite field L to perform computations and uses
the finite fields procedure of [PR11], which roughly takes an expected number
of O(d3Y d2X + d2Y dX t0 log p) arithmetic operations in L to compute the singular parts of all the rational Puiseux expansions above 0. Here t0 := [L : Fp ]
denotes the degree of the field extension L/Fp and dX and dY are the degree
of F in the variable X and Y respectively.
Nevertheless, as we are only interested in particular cases where the singular parts can be immediately generated (see Subsections 4.4.3 and 4.4.4),
we are not going to use this procedure.
4.4.
Examples
In this section we apply our algorithmic method in order to compute a
geometric solution of certain zero–dimensional polynomial equation systems.
In Section 4.4.1 we treat the case of Pham–Brieskorn systems. In Section
4.4.2 we treat a family of systems which arise from a semidiscretization of
certain parabolic differential equations with nonlinear source terms and nonlinear boundary conditions. Finally, in Section 4.4.4 we treat a generalization
of Reimer systems, which we called generalized Reimer systems.
In all the above cases, we “deform” the polynomial equation system under
consideration to a one–dimensional polynomial equation system satisfying
the hypotheses of Theorem 4.7. Then the algorithm underlying the proof of
Theorem 4.12 yields an efficient procedure to compute a geometric solution
of the original zero–dimensional polynomial equation system.
We observe that all these examples are particular instances of a generalized Pham system. Nevertheless, the deformations we shall introduce have a
ramified fibre with a very simple infinitesimal structure. This will allow us to
slightly improve the cost that we obtain when applying the main algorithm
of Chapter 3 to the polynomial systems which arise in these examples.
107
4.4. EXAMPLES
4.4.1.
Pham–Brieskorn systems
Let us fix n, d ∈ N. Let g1 , . . . , gn ∈ Q[X] := Q[X1 , . . . , Xn ] satisfy
deg(gi ) < d and gi (0, . . . , 0) 6= 0 for 1 ≤ i ≤ n. Let us define f1 , . . . , fn ∈
Q[X] by:
f1 := X1d − g1 , . . . , f1 := Xnd − gn .
(4.29)
A system of this form is called a Pham–Brieskorn system (see e.g. [LV98],
[GV98], [Bom00], [PS04]). It is easy to see that f1 , . . . , fn form a regular
sequence of Q[X].Therefore, f1 , . . . , fn define a zero–dimensional affine subvariety Ve of An . Our aim is to compute a geometric solution of this variety
Ve .
Let E be an indeterminate over Q and define F1 , . . . , Fn ∈ Q[E, X] by:
F1 := X1d − Eg1 , . . . , Fn := Xnd − Egn .
(4.30)
Let V be the affine subvariety of An+1 defined by the polynomials F1 , . . . , Fn ,
and let π : V → A1 be the morphism defined by π(ε, x) := ε. We observe
that π −1 (1) = {1} × Ve and π −1 (0) = {0} ⊂ An+1 hold.
In Section 4.4.3 we exhibit an algorithm which computes a geometric solution of the variety V . Furthermore, specializing the polynomials of Q[E, X]
which constitute this geometric solution into the value E = 1 we shall obtain
a geometric solution of Ve .
4.4.2.
Systems coming from a semidiscretization of certain parabolic differential equations
In this section we consider a family of polynomial equation systems which
arises in the analysis of the stationary solutions of a numerical approximation,
obtained by a semidiscretization in space, of certain parabolic differential
equations with nonlinear source terms and nonlinear boundary conditions
(see e.g. [BR01], [FGR02]).
Let us fix n, d ∈ N with d ≥ 2. Let T be an indeterminate over Q, and
let g, h ∈ Q[T ] \ {0} satisfy deg(g) < d and deg(h) = d. Let us write h =
aT d + h1 (T ) with a 6= 0 and deg(h1 ) < d. Let f1 , . . . , fn be the polynomials
108
of Q[X] := Q[X1 , . . . , Xn ] defined in the following way:
f1 := 2(n − 1)2 (X2d − X1d ) − g(X1 ),
d
d
fi := (n − 1)2 (Xi+1
− 2Xid + Xi−1
) − g(Xi ), (2 ≤ i ≤ n − 1)
2
d
d
fn := 2(n − 1) (Xn−1 − Xn ) + 2(n − 1)h(Xn ) − g(Xn ).
(4.31)
An important case of study is that of the stationary solutions of the porous
medium equation with nonlinear source terms and nonlinear boundary condition (see e.g. [Hen81], [CFQ91b]). Typical discretizations of this problem
lead for example to instances of system (4.31) with h := T d and g := T (see
e.g. [FGR02]).
Let Ve be the affine subvariety of An defined by the polynomials f1 , . . . , fn .
Our aim is to exhibit an efficient algorithm which computes a geometric
solution of the variety Ve . For this purpose, let f := (f1 , . . . , fn ), en :=
(0, . . . , 0, 1) ∈ Qn , G := (g(X1 ), . . . , g(Xn )), and X d := (X1d , . . . , Xnd ). Let
A ∈ Qn×n be the following nonsingular tridiagonal matrix:

−2
2
 1 −2 1

2
.. .. ..
A := (n − 1) 
.
.
.


1 −2 1
2 −2 +

2a
n−1
Then the polynomials f1 , . . . , fn can be expressed as:
f t = A · (X d )t + 2(n − 1)h1 (Xn )etn − Gt ,



.


(4.32)
where t denotes transposition.
In order to solve the system defined by the polynomials in (4.32), we
introduce a new indeterminate E and consider the following polynomials of
Q[E, X]:
(Fe1 , . . . , Fen ) t := A·(X d )t+E 2(n−1)h1 (Xn )ent −G t −2(n−1)E(1−E)ent . (4.33)
Let V be the affine subvariety of An+1 defined by the polynomials Fe1 , . . . , Fen
and let π : V → A1 be the morphism defined by π(ε, x) = ε. We observe
that π −1 (1) = {1} × Ve and π −1 (0) = {0} ⊂ An+1 . Since the matrix A is
109
4.4. EXAMPLES
nonsingular, multiplying both sides of (4.33) by A−1 we obtain the following
polynomials, whose zero set also defines the variety V :
(F1 , . . . , Fn )t := (X d )t + EA−1 2(n − 1)h1 (Xn )etn − Gt − E(E − 1)v t ,
(4.34)
n−1
where v := 2a (1, . . . , 1). In Section 4.4.3 we exhibit an algorithm computing
a geometric solution of the variety V . By specializing the polynomials of
Q[E, X] which constitute this geometric solution into the value E = 1 we
shall obtain a geometric solution of our input variety Ve .
4.4.3.
A common approach to both examples
In this section we describe an algorithm which finds a geometric solution
of the variety defined by any system of the form (4.30) and (4.34). Then,
we shall specialize the polynomials of Q[E, X] which form such geometric
solution into the value E = 1 in order to obtain a geometric solution of the
variety defined by the corresponding system of the form (4.29) and (4.31).
Let us fix n, d ∈ N. For 1 ≤ i ≤ n, let Hi ∈ Q[E, X] satisfy deg Hi ≤ d − 1
and αi := Hi (0, 0) 6= 0. Suppose further that we are given a straight–line program computing the polynomials H1 , . . . , Hn using T arithmetic operations
in Q.
For 1 ≤ i ≤ n, let us define Fi ∈ Q[E, X] by the following expression:
Fi := Xid − EHi (E, X).
(4.35)
Let I be the ideal of Q[E, X] generated by F1 , . . . , Fn and let V be the affine
subvariety of An+1 defined by I. Let π : V → A1 denote the restriction to
V of the canonical projection onto the first coordinate. Our purpose is to
compute a geometric solution of {1} × Ve := π −1 (1).
It is easy to see that any system of the form (4.30) and (4.34) is a particular instance of a system of the form (4.35). In order to apply our algorithmic
method, we first show in Lemmas 4.13 and 4.14 below that the polynomials
F1 , . . . , Fn of (4.35) form a regular sequence of Q[E, X], the ideal I ⊂ Q[E, X]
they generate is radical, and the morphism π is finite and generically unramified.
Lemma 4.13. The polynomials F1 , . . . , Fn form a regular sequence of Q[E, X]
and the morphism π is finite.
110
Proof. From Buchberger’s first criterion (see e.g. [BW93]), we conclude that
for 1 ≤ i ≤ n the polynomials F1 , . . . , Fi form a Gröbner basis of the ideal
they generate with respect to the graded lexicographical order induced by
the ordering X1 > · · · > Xn > E. This implies that the affine variety of An+1
defined by F1 , . . . , Fi has codimension i for 1 ≤ i ≤ n. Then F1 , . . . , Fn form
a regular sequence of Q[E, X].
Furthermore, we observe that the leading monomial of Fi under this order
is Xid for 1 ≤ i ≤ n. Therefore, the set {X1i1 · · · Xnin : 0 ≤ i1 , . . . , in <
d} is a basis as Q[E]–module of Q[E, X]/I. This proves that π is a finite
morphism.
For 1 ≤ i ≤ n, let Gi ∈ Q[E, X] be the following polynomial:
Gi (E, X) := E −d Fi (E d , EX).
f ⊂ An+1 be the affine variety defined by G1 , . . . , Gn , and let π
Let W
e :
1
f
W → A be the morphism induced by the canonical projection onto the
first coordinate. We claim the morphism π
e is generically unramified.
Let us observe that for ε 6= 0 we have #(e
π −1 (ε)) = #(π −1 (εd )). Therefore,
from the fact that the morphism π is finite we easily conclude that π
e is
f
dominant and dim W ≥ 1 holds. Furthermore, from the fact that Q(V )
f ) is also a zero–
is a zero–dimensional Q(E)–algebra, we deduce that Q(W
f is a one–dimensional variety.
dimensional Q(E)–algebra. This shows that W
Let us fix ε ∈ A1 . Taking into account that degX Gi (ε, X) = d for
1 ≤ i ≤ n, from the Bézout inequality (2.1) we deduce that deg π
e−1 (ε) ≤ dn
holds. On the other hand, for 1 ≤ i ≤ n we have Gi (0, X) = Xid − αi , where
αi = Hi (0, 0) 6= 0. This implies that π
e−1 (0) has cardinality dn . We conclude
−1
that any generic fibre π
e (ε) has cardinality dn .
Lemma 4.14. I is a radical ideal and the morphism π is generically unramified.
Proof. For a generic choice ε ∈ A1 , we have #(e
π −1 (ε)) = #(π −1 (εd )) = dn .
−1
This implies that there exists a fibre π (ε) of cardinality dn . On the other
hand, applying the Bézout inequality (2.1) we see that #(π −1 (ε)) ≤ dn holds
for any ε ∈ A1 . We conclude that #(π −1 (ε)) = dn holds for any generic
choice of the value ε ∈ A1 .
4.4. EXAMPLES
111
Let ε be a generic element of A1 . Then
dimC C[X]/ F1 (ε, X), . . . , Fn (ε, X) = dn = deg π −1 (ε).
This implies (see e.g. [CLO98, Corollary 2.6]) that π −1 (ε) is a smooth variety and the polynomials F1 (ε, X), . . . , Fn (ε, X) generate a radical ideal of
C[X]. In particular, we have that the Jacobian determinant JF (ε, X) :=
det(∂Fi /∂Xj )1≤i,j≤n (ε, X) does not vanish on any point x ∈ An with (ε, x) ∈
π −1 (ε). Thus, JF (E, X) is not a zero divisor of Q[E, X]/I and π is generically
unramified. Finally, since F1 , . . . , Fn form a regular sequence of Q[E, X], from
[Eis95, Theorem 18.15] we deduce that the ideal I is radical.
Let us observe that the origin 0 ∈ An+1 is the only point of π −1 (0).
Therefore, there are deg(π) = dn branches of the curve V passing through
0 ∈ An+1 .
For F ∈ Q[E, X], let us write F (E d , EX) = E α f (X) + O(E α+1 ), with
f 6= 0. We define the initial term of F with respect to the weight (d, 1, . . . , 1)
as the polynomial ind (F ) := f . Let ind (I) ⊂ Q[X] be the ideal generated
by the set {ind (F ) : F ∈ I} and let W ⊂ An be the affine variety defined by
ind (I).
Lemma 4.15. W = V (X1d −α1 , . . . , Xnd −αn ) and G1 , . . . , Gn form a standard
basis.
Proof. Let us observe that the set {ind (F ) : F ∈ I} is contained in the set
of initial terms (in the sense of Section 4.2) of the polynomials of the ideal
(G1 , . . . , Gn ). Let F ∈ (G1 , . . . , Gn ), and let us write F = E α Fe(E, X), with
α ≥ 0 and Fe(0, X) 6= 0. Since E is not a zero divisor of the Q–algebra
Q[E, X]/(G1 , . . . , Gn ), we conclude that Fe ∈ (G1 , . . . , Gn ) holds. Then
in(Fe) = Fe(0, X) ∈ G1 (0, X), . . . , Gn (0, X) = (X1d − α1 , . . . , Xnd − αn ),
which implies that ind (I) ⊂ (X1d − α1 , . . . , Xnd − αn ) holds and G1 , . . . , Gn
form a standard basis. On the other hand,
(X1d − α1 , . . . , Xnd − αn ) = ind (F1 ), . . . , ind (Fn ) ⊂ ind (I),
from which the statement of Lemma 4.15 follows.
112
Since there are dn branches of V lying above 0 and deg W = dn , we
conclude that the system of (classical) Puiseux expansions of the branches
of the curve V lying above 0 has regularity index 1, and the singular parts
of its expansions are represented by the points of W .
Lemmas 4.14 and 4.15 show that the polynomials of (4.35) satisfy the
hypotheses of Theorems 4.7 and 4.12. In order to apply the algorithm underlying Theorem 4.12 to our input system, we first need an explicit description
of the set of singular parts of a system of rational Puiseux expansions of the
branches of V lying above 0. For this purpose, we observe that the set of
singular parts is given by
d j1 1/d
(T , ξ α1 T, . . . , ξ jn αn1/d T ); 0 ≤ j1 , . . . , jn < d ⊂ Q[T ]n+1 ,
1/d
1/d
where ξ ∈ Q is a primitive d–th root of 1 and α1 , . . . , αn ∈ Q are d–
−1/d
th roots of α1 , . . . , αn respectively. Replacing T by α1 T we obtain the
following system of rational Puiseux expansions of the branches of V lying
above 0:
−1 d
1/d
(α1 T , T, ξ j2 β2 T, . . . , ξ jn βn1/d T ); 0 ≤ j2 , . . . , jn < d ⊂ Q[T ]n+1 ,
1/d
1/d
where β2 , . . . , βn ∈ Q are d–th roots of β2 := α1−1 α2 , . . . , βn := α1−1 αn
respectively. With the notations of Section 4.1, we have g = 1, e1 = d,
f1 = dn−1 .
Let Y2 , . . . , Yn be new indeterminates over Q. Let
1/d
W0 := (ξ j2 β2 , . . . , ξ jn βn1/d ); 0 ≤ j2 , . . . , jn < d = V (Y2d − β2 , . . . , Ynd − βn ).
Then we see that a geometric solution of the variety W0 yields the polynomi(1)
(1)
als q (1) , f2 , . . . , fn required for the application of the algorithm of Theorem
4.12.
Let U := γ2 Y2 + · · · + γn Yn be a linear form of Q[Y2 , . . . , Yn ] inducing a
primitive element of the Q–algebra extension Q ֒→ Q[W0 ]. In order to compute a geometric solution of W0 we apply the algorithm underlying the proof
of Lemma 3.11. Fix ρ ≥ 2 and suppose that the coefficients of γ2 , . . . , γn are
randomly chosen in the set {1, . . . , 4nρd3n−3 }. By Lemma 3.11 and Proposition 3.14 we conclude that such a geometric solution can be computed with
O(M(d2n−2 ) arithmetic operations in Q. Finally, applying Theorem 4.12 we
obtain the following result.
113
4.4. EXAMPLES
Theorem 4.16. There exists a computation tree computing a geometric solution of the variety V with O(T M(dn )2 ) arithmetic operations in Q.
The geometric solution provided by Theorem 4.16 consists of a randomly
chosen linear form U ∈ Q[X] and polynomials mu , v1 , . . . , vn ∈ Q[E, Z].
Suppose that U is also a primitive element of the original variety {1} × Ve =
V ∩ ({1} × An ). Specializing mu , v1 , . . . , vn into the value E = 1, we obtain
polynomials mu (1, Z), v1 (1, Z), . . . , vn (1, Z) of Q[X] defining a (eventually
non–reduced) Shape–Lemma–like representation of Ve . Therefore, computing
a square–free representation of mu (1, Z), and cleaning the multiple factors of
the polynomial mu (1, Z) out of v1 (1, Z), . . . , vn (1, Z) we obtain a geometric
solution of Ve with the same complexity estimate (see [GLS01] for details).
This result improves the rough O(3n d2n ) complexity estimate of [MP97].
Let us also mention the results of [MP00], where the authors announce a
rough O(d2n ) complexity estimate for approximating one root of a Pham
system. Comparing our result with the rough O(T d2n−1 ) complexity estimate provided by the application of the algorithm of [GLS01] to this case,
we see that the performance of [GLS01] is better. Nevertheless, let us observe that the leading term d2n of our complexity estimate can be expressed
as δ degE mu and we are dealing in this case with an “ill-conditioned” system, for which the worst case estimates δ = dn and degE mu = dn hold. If
the input system satisfies degE mu ≪ dn , then the performance of [GLS01]
does not change, whereas in our complexity estimate the d2n factor reduces
accordingly. Furthermore, if degE mu = 1, we achieve the lower bound dn of
this factor (see [CGH+ 03]).
4.4.4.
Reimer Systems
In this section we consider another family of examples called (generalized)
Reimer systems (compare [BM96]). Let us fix n ∈ N, and let us define
f1 , . . . , fn ∈ Q[X] := Q[X1 , . . . , Xn ] in the following way:
fi := αi +
n
X
ai,j Xji+1 ,
(4.36)
j=1
where ai,j , αi (1 ≤ i, j ≤ n) are generic elements of Q (see Lemma 4.17
below) with αi , ai,i 6= 0 for 1 ≤ i ≤ n. Let Ve be the affine subvariety of An
114
defined by f1 , . . . , fn . Our purpose is to compute a geometric solution of Ve .
Our next result shows that Ve has dimension zero and degree (n + 1)!.
Lemma 4.17. Let U := (Ui,j )1≤i,j≤n be a matrix of indeterminates and let
H1 , . . . , Hn be the elements of Q[U, X] defined in the following way:
Hi := αi +
n
X
Ui,j Xji+1 .
j=1
2
Then there exists a non–empty Zariski open set U ⊂ An with the following
property: for any u ∈ U, the affine subvariety of An defined by the polynomials H1 (u, X), . . . , Hn (u, X) has dimension 0 and degree (n + 1)!.
2
Proof. Let Z be the affine variety of An +n defined by H1 , . . . , Hn and let
2
πU : Z → An be the morphism defined by π(u, x) = u. Let ℘ be the prime
ideal of Q[U ] generated by the set {Ui,j ; 1 ≤ i, j ≤ n, i 6= j}. We claim that
H1 , . . . , Hn form a regular sequence of Q[U ]℘ [X].
In order to prove this claim, following [HJS+ 02], we define a “triangular”
(i)
sequence (Rj )1≤i≤n,i+1≤j≤n ⊂ Q[U, X] in the following way:
(1)
Rj := ResX1 (H1 , Hj ) for j = 2, . . . , n.
(i)
(i−1)
Rj := ResXi (Ri
(i−1)
, Rj
) for 2 ≤ i ≤ n − 1 and i + 1 ≤ j ≤ n.
(i−1)
From elementary properties of the resultant we see that Ri
is a nonzero
(i−1)
(i−1)
element of Q[U, Xi , . . . , Xn ] ∩ (H1 , . . . , Hi ), with degX Ri
= degXi Ri
.
Furthermore, a recursive argument shows that the coefficient of the highest
(i−1)
power of Xi occurring in Ri
does not belong to the prime ideal ℘. We
conclude that H1 , . . . , Hi define an ideal of Q[U ]℘ [X] of Krull dimension n−i.
This implies that H1 , . . . , Hn form a regular sequence of Q[U ]℘ [X].
(n−1)
Furthermore, the polynomial Rn
gives an integral dependence equation for the coordinate class of Xn in the ring Q[U ]℘ [X1 , . . . , Xn ]/(H1 , . . . , Hn )
(i−1)
over the ring Q[U ]℘. Then a recursive argument with the polynomials Ri
for 1 ≤ i ≤ n shows that
Q[U ]℘ ֒→ Q[U ]℘ [X1 , . . . , Xn ]/(H1 , . . . , Hn )
(4.37)
115
4.4. EXAMPLES
is an integral Q–algebra extension.
2
We conclude that there exists a Zariski neighborhood Ue ⊂ An of V (℘)
such that πU |Z∩(Ue×An ) : Z∩(Ue×An ) → Ue is a finite morphism and Z∩(Ue×An )
is an equidimensional variety of dimension n2 . This shows that for any choice
of u ∈ Ue the variety Z ∩ {U = u} = πU−1 (u) has dimension 0.
Now we show that the existence of the Zariski open set U ⊂ Ue of the
statement of the lemma. First, we observe that the Bézout inequality (2.1)
implies deg(πU−1 (u)) ≤ (n + 1)! for any u ∈ Ue. On the other hand, for any
nonsingular diagonal matrix u(0) ∈ Ue we have deg(πU−1 (u(0) )) = (n + 1)!. We
conclude that there exists a non–empty Zariski open set U ⊂ Ue such that
deg(πU−1 )(u) = (n + 1)! holds for any u ∈ U.
Let us observe that for any u ∈ U we have that C[X]/(H1 (u,X), . . . , Hn (u,X))
is a finite–dimensional C–vector space of dimension at most (n + 1)!. On the
other hand, we have #(πU−1 (u)) = (n+1)!. We conclude that the polynomials
H1 (u, X), . . . , Hn (u, X) generate a radical zero–dimensional ideal of C[X],
and hence the Jacobian determinant JH (u, X) := det(∂Hi /∂Xj )1≤i,j≤n (u, X)
does not vanish on any point x with (u, x) ∈ πU−1 (u). This implies that JH
does not vanish on any point of Z ∩ (U × An ).
In order to solve a system of the form (4.36) with a := (ai,j )1≤i,j≤n ∈ U ,
let us introduce an indeterminate E over Q and the following elements of
Q[E, X]:
Fi := αi E i+1 + ai,i Xii+1 +
X
ai,j EXji+1
(1 ≤ i ≤ n).
(4.38)
1≤j≤n
j6=i
Let V be the affine subvariety of An+1 defined by F1 , . . . , Fn and let π : V →
A1 be the morphism defined by π(ε, x) := ε. We have π −1 (1) = {1} × Ve and
π −1 (0) = {0} ⊂ An+1 . We are going to show that F1 , . . . , Fn form a regular
sequence of Q[E](E) [X] and generate a radical ideal of Q[E](E) [X], and the
morphism π is dominant and generically unramified.
For this purpose, let us define G1 , . . . , Gn ∈ Q[E, X] in the following way:
Gi := E −(i+1) Fi (E, EX) = αi +
n
X
j=1
gi,j Xji+1 ,
116
f be the affine subvariety
where gi,j := ai,j E for i 6= j and gi,i := ai,i . Let W
n+1
f → A1 be the morphism defined
of A
defined by G1 , . . . , Gn , and let π
e:W
2
by π
e(ε, x) = ε. Observe that g(1) ∈ U holds, where U ⊂ An is the Zariski
open set of the statement of Lemma 4.17. Therefore, for a generic choice
ε ∈ A1 , we have g(ε) ∈ U. Taking into account the remarks after the proof
of Lemma 4.17, we conclude that π
e is dominant and generically unramified.
Finally, since #(e
π −1 (ε)) = #(π −1 (ε)) holds for any ε 6= 0, we deduce the
following result.
Lemma 4.18. The morphism π is dominant and generically unramified.
On the other hand, we have the following result.
Lemma 4.19. F1 , . . . , Fn form a regular sequence in Q[E](E) [X] and generate
a radical ideal of Q[E](E) [X].
Proof. For 1 ≤ i ≤ n, let Fbi ∈ [E, X0 , . . . , Xn ] denote the homogenization of
the polynomial Fi with respect to the variables X. We have Fbi ≡ ai,i Xii+1
mod (E). Following [HJS+ 02], we define the following “triangular” sequence
b(i) )1≤i≤n,i+1≤j≤n of Q[E, X]:
(R
j
b(1) := ResX1 (Fb1 , Fbj ) for j = 2, . . . , n.
R
j
b(i) := ResX (R
b(i−1) , R
b(i−1) ) for 2 ≤ i ≤ n − 1 and i + 1 ≤ j ≤ n.
R
i
j
i
j
b(i) is an hoFrom the elementary properties of the resultant we deduce that R
j
b
b
mogeneous polynomial of (F1 , . . . , Fj )∩Q[E, X0 , Xi+1 , . . . , Xn ]. Furthermore,
taking into account the congruence relation Fbi ≡ ai,i Xii+1 mod (E), a simple
b(i−1) ≡ ci X mi mod (E) holds for suitable
recursive argument shows that R
i
i
(i−1)
ci ∈ Q \ {0} and mi ∈ N. This shows that the coefficient of Ximi in Ri
does not belong to the prime ideal (E) ⊂ Q[E]. Specializing the variable X0
into the value X0 = 1, with a similar argument as in the proof of Lemma
4.17 we conclude that F1 , . . . , Fn form a regular sequence of Q[E](E) [X] and
Q[E](E) ֒→ Q[E](E) [X]/(F1 , . . . , Fn )
is an integral Q–algebra extension.
4.4. EXAMPLES
117
Finally, since F1 , . . . , Fn form a regular sequence of Q[E](E) [X], and the
morphism π is generically unramified, applying [Eis95, Theorem 18.15] as in
Lemma 4.14 we conclude that the ideal generated by F1 , . . . , Fn in Q[E](E) [X]
is radical.
Let us observe that the origin 0 ∈ An+1 is the only point of π −1 (0).
Therefore, there are deg(π) = (n + 1)! branches of V passing through 0 ∈
Cn+1 .
For any F ∈ Q[E](E) [X], let us write F (E, EX) = E α Fe(E, X) with Fe ∈
Q[E](E) [X]\(E)Q[E](E) [X]. We define the initial term of F with respect to the
weight (1, . . . , 1) as in1 (F ) := Fe(0, X). Let I be the ideal of Q[E](E) [X] generated by F1 , . . . , Fn , and let in1 (I) ⊂ Q[X] be the ideal generated by the set
{in1 (F ) : F ∈ I}. Let W := V (in1 (I)) ⊂ An .
Lemma 4.20. W = V (a1,1 X12 − α1 , . . . , an,n Xnn+1 − αn ) and G1 , . . . , Gn form
a standard basis in Q[E](E) [X].
Proof. Let us observe that the set {in1 (F ) : F ∈ I} is contained in the set
of initial terms (in the sense of Section 4.2) of the polynomials of the ideal
(G1 , . . . , Gn ) ⊂ Q[E](E) [X]. Let F ∈ (G1 , . . . , Gn ) and write F = E α Fe(E, X),
with α ≥ 0 and Fe(0, X) 6= 0. Since E is not a zero divisor of the Q–algebra
Q[E](E) [X]/(G1 , . . . , Gn ), we conclude that Fe ∈ (G1 , . . . , Gn ) holds. Then
in1 (Fe) = Fe(0,X) ∈ G1 (0,X), . . . , Gn (0,X) = (a1,1 X12 −α1 , . . . , an,n Xnn+1−αn ),
which implies that in1 (I) ⊂ (a1,1 X12 − α1 , . . . , an,n Xnn+1 − αn ) holds and
G1 , . . . , Gn form a standard basis. On the other hand, we have the inclusion
(a1,1 X12 − α1 , . . . , an,n Xnn+1 − αn ) = in1 (F1 ), . . . , in1 (Fn ) ⊂ in1 (I),
from which the lemma follows.
Since there are (n + 1)! branches of V lying above 0 and deg W = (n + 1)!,
we conclude that the system of (classical) Puiseux expansions of the branches
of the curve V lying above 0 has regularity index 1, and the singular parts
of its expansions are represented by the points of W .
Lemmas 4.18, 4.19 and 4.20 show that the polynomials F1 , . . . , Fn of
(4.38) satisfy all the hypotheses of Theorems 4.7 and 4.12 (see the remark
118
right before Section 4.2.1). In order to apply the algorithm underlying Theorem 4.12 we need a description of the singular parts of the branches of V
lying above 0. A similar argument as in Section 4.4.3 shows that, with the
notations of Section 4.1, g = 1, e1 = 1 and f1 = (n + 1)! in this case. Hence,
we have that a geometric solution of the variety W yields the polynomials
(1)
(1)
q (1) , f1 , . . . , fn required for the application of Theorem 4.12. Such a geometric solution can be obtained in O(nM((n + 1)!)2 ) arithmetic operations
in Q, using a similar algorithm to that of Section 4.4.3. Finally, applying
Theorem 4.12 we obtain the following result.
Theorem 4.21. There exists a straight–line program computing a geometric
solution of the variety V in O(nM((n + 1)!)2 ) arithmetic operations in Q.
In order to obtain a geometric solution of the variety {1} × Ve = π −1 (1)
from the geometric solution of V provided by Theorem 4.21, we proceed in
a similar way as in Section 4.4.3 (see the remarks after Theorem 4.16).
Chapter 5
Deformation techniques for sparse
systems
This chapter deals with the symbolic computation of all the solutions
to zero-dimensional sparse multivariate polynomial equation systems, i.e.,
systems with a finite number of common complex zeros defined by a sparse
square system of polynomials. It is based on an article of the same title that
I co-authored with Gabriela Jerónimo, Guillermo Matera, and Pablo Solernó
([JMSW09]).
The origins of sparse elimination theory can be traced back to the results by D.N. Bernstein, A.G. Kushnirenko and A.G. Khovanski ([Ber75],
[Kus76], [Kho78]) that bound the number of solutions of a polynomial system in terms of a combinatorial invariant associated to the set of exponents
of the monomials arising with nonzero coefficients in the defining polynomials. More precisely, the Bernstein-Kushnirenko-Khovanski (BKK for short)
theorem asserts that the number of isolated solutions in the n-dimensional
complex torus (C∗ )n of a polynomial system of n equations in n unknowns is
bounded by the mixed volume of the family of Newton polytopes of the corresponding polynomials. Hence sparse elimination looks for techniques that
profit from a low mixed volume of the underlying system or other sparsity
parameters.
Numeric (homotopy continuation) methods for sparse systems are typically based on a specific family of deformations called polyhedral homotopies
([HS95], [VVC94], [VGC96], [Roj03]). Polyhedral homotopies preserve the
119
120
CHAPTER 5. DEFORMATIONS FOR SPARSE SYSTEMS
Newton polytope of the input polynomials and yield an effective version of
the BKK theorem (see, e.g., [HS95], [HS97]).
In this chapter we combine the homotopic procedures of [HS95] with
the symbolic deformation techniques developed in the previous chapters in
order to derive a symbolic probabilistic algorithm for solving sparse zerodimensional polynomial systems with cubic cost in the size of the combinatorial structure of the input system. Our main result may be stated as
follows (see Theorem 5.23 below for a precise statement).
Theorem 5.1. Let f1 , . . . , fn be polynomials in Q[X1 , . . . , Xn ] such that the
system f1 = 0, . . . , fn = 0 defines a zero-dimensional affine subvariety V
of Cn . Denote by ∆1 , . . . , ∆n ⊂ Zn≥0 the supports of f1 , . . . , fn , and assume
that 0 ∈ ∆i for 1 ≤ i ≤ n and the mixed volume D of the Newton polytopes
Q1 := Conv(∆1 ), . . . , Qn := Conv(∆n ) is nonzero.
Then, we can probabilistically compute a geometric solution of the va′
riety
in
P Q, with N :=
P O(N DD ) arithmetic operations
P V using roughly
′
1≤i≤n M(∆, Q1 ,
1≤i≤n M(∆, Q1 , . . . , Qn ) and D :=
1≤i≤n #∆i , D :=
. . . , Qi−1 , Qi+1 , . . . , Qn ), where ∆ denotes the standard n-dimensional simplex and M stands for mixed volume.
As input we are given polynomials f1 , . . . , fn ∈ Q[X1 , . . . , Xn ] which define a zero-dimensional variety of An (C). We assume that the combinatorics
of the polyhedral deformation mentioned above are known. More precisely,
we assume that we are given a certain collection of subsets of the input
supports ∆1 , . . . , ∆n , which defines a fine-mixed subdivision of ∆1 , . . . , ∆n ,
together with the lifting function which yields such a subdivision (for precise definitions see [HS95, Section 2] or Section 5.1 below). For an efficient
algorithm computing these objects see, for instance, [LL01].
The input of our algorithm is the standard sparse representation of the
polynomials f1 , . . . , fn ∈ Q[X1 , . . . , Xn ], that is, the list of exponents of all
nonzero monomials arising in f1 , . . . , fn together with the corresponding coefficients. We observe that in our setting there are no significant differences
between the sparse and the straight–line program representation. Indeed,
any polynomial f ∈ Q[X1 , . . . , Xn ] of degree at most d > 0 having support
∆ ⊂ Zn≥0 can be evaluated with O(n#∆ log d) arithmetic operations in Q.
In this sense, we see that f has a straight–line program representation whose
size is of the same order as its standard sparse representation, and can be
121
efficiently obtained from the latter. On the other hand, from a straight–line
program which evaluates a polynomial f ∈ Q[X1 , . . . , Xn ] of (known) support ∆ with L arithmetic operations, the corresponding sparse representation
can be easily obtained by a process of multipoint evaluation and interpolation with cost O(L#∆), up to logarithmic terms. Since the routines of our
procedure are of black-box type (cf. [CGH+ 03]), that is, they only call the
input polynomials and their first derivatives for substitutions of the variables
X1 , . . . , Xn into values belonging to suitable commutative zero-dimensional
algebras, we conclude that the straight–line program representation of intermediate results is better suited than the sparse one. In particular, we note
that computing the first derivatives of a multivariate polynomial can be done
more efficiently for polynomials given by straight–line programs than by their
sparse encoding (cf. [BS83]).
The complexity of our algorithm is mainly expressed in terms of three
quantities which measure the size of the combinatorial
P structure of the input
system: the number of nonzero coefficients P
N := 1≤i≤n #∆i and the mixed
′
volumes D := M(Q1 , . . . , Qn ) and D = 1≤i≤n M(∆, Q1 , . . . , Qi−1 , Qi+1 ,
. . . , Qn ). While D represents the (optimal) number of paths which are followed during our homotopy, the quantity D′ is an arithmetic analogue of D
(see [PS03], [PS05]) which measures the “precision” at which the paths of
our homotopy must be followed. We observe that the invariant D′ is also
optimal for a generic choice of the coefficients of the polynomials f1 , . . . , fn
(see Lemma 5.3 below; compare also with [PS08, Theorem 1.1]). Therefore,
we may paraphrase our complexity estimate as saying that it is cubic in the
combinatorial structure of the input system, with a geometric component,
an arithmetic component and a component related to the size of the input
data. In this sense, we see that the cost of our algorithm strongly resembles the cost O(N Dµ2 ) of numerical continuation algorithms, where µ is the
highest sparse condition number arising from the application of the Implicit
Function Theorem to the points of the paths which are followed (cf. [Ded97];
see also [MR04] for a probability analysis of the condition numbers of sparse
systems).
Our result improves and refines the estimate of Chapter 4 in the case of a
sparse system, which is expressed as a fourth power of D and the maximum
of the degrees of two varieties associated with the input (we observe that this
maximum is an upper bound for the parameter D′ ). On the other hand, it
also improves [Roj99], [Roj00], which solve a sparse system with a complexity
122
which is roughly quartic in the size of the combinatorial structure of the input
system. We shall also provide explicit estimates of the error probability of
all the steps of our algorithm. This might be seen as a further contribution
to the symbolic stage of the probabilistic semi-numeric method of [HS95],
which lacks such an analysis of error probability.
The interest in tropical algebraic geometry has spawned results that generalise ours. Herrero et al. ([HJS13]) extend our results to the case where the
square system f1 = 0, . . . , fn = 0 has a positive-dimensional solution set. In
fact, the article (op. cit) describes an algorithm for computing one representative point in each irreducible component. More precisely, it outputs a list of
geometric solutions —one per equidimensional component— where each geometric solution describes one point per irreducible component. The algorithm
2
˜ 4
uses O
P(nn dN D ) arithmetic operations in Q, where d := max1≤j≤n {deg(fj )},
N = j=1 #(Aj ∪ ∆) and D = M Vn (A1 ∪ ∆, . . . , An ∪ ∆).
Likewise, [Ver09] presents a different method for representing all the solutions of a possibly positive-dimensional polynomial system having at least
as many equations as unknowns. It may also be worthwhile to mention that
other authors of polyhedral homotopy methods aim to solve the problem by
solving the face enumeration problem of finding all the lower facets of a mixed
subdivision (see, e.g., [MTK07], [Miz08] and the bibliography therein).
The main algorithm of this chapter proceeds in two main steps: in the
first step, the polyhedral deformation introduced in [HS95] is applied to solve
an auxiliary generic sparse system with the same combinatorial structure as
the input polynomials (Section 5.2; see also Section 5.2.1 for a discussion
on the genericity conditions underlying the choice of the coefficients of the
corresponding polynomials). In the second step the solutions of this generic
system enable us to recover the solutions of the given system by means of a
standard homotopic deformation (see Section 5.3).
In the first step, we partake to solve a system h1 = 0, . . . , hn = 0 with the
same supports ∆1 , . . . , ∆n as f1 , . . . , fn but with generic coefficients, which
are chosen randomly. In order to do this, the polyhedral homotopy of [HS95]
introduces a new variable T and deforms each polynomial hi by multiplying
each nonzero monomial of hi by a power of T (which is determined by the
given lifting function). The roots of the resulting parametric system are algebraic functions of the parameter T whose expansions as Puiseux series can be
obtained by “lifting” the solutions from certain associated zero-dimensional
5.1. SPARSE ELIMINATION
123
polynomial systems, which in turn can be easily solved due to their specific
structure (see Section 5.2.4 for details). This enables us to compute a geometric solution of the zero set of this parametric system (Section 5.2.5).
Substituting 1 for T in the computed polynomials we obtain a geometric
solution of the set of common zeros of h1 , . . . , hn (Section 5.2.6).
For the sake of comprehensiveness, throughout Section 5.2 the whole first
step of the algorithm will be illustrated with a bivariate polynomial example
borrowed from [HS95, Example 2.7].
After solving the system h1 = 0, . . . , hn = 0, in the second step the solutions to the input system f1 = 0, . . . , fn = 0 are recovered by considering a
second homotopy of type T f1 + (1 − T )h1 , . . . , T fn + (1 − T )hn (see Section
5.3). As in the first step, the algorithm first solves this parametric system
(Section 5.3.1) and then, substituting 1 for T , a complete representation of
the solution set of the input system is obtained. This representation eventually includes multiplicities, which are removed in a further computation
(Section 5.3.2).
5.1.
Sparse Elimination
Here we introduce some notions and notations of convex geometry and
sparse elimination theory (see, e.g., [GKZ94], [HS95], [Roj03]) that will be
used in the sequel.
Let X1 , . . . , Xn be indeterminates over Q and write X := (X1 , . . . , Xn ).
n
For
X q := X1q1 · · · Xnqn . Let f :=
P q :=q (q1 , . . . , qn ) ∈ Z , we use the notation
−1
] := Q[X1 ,X1−1, . . . ,Xn ,Xn−1 ].
q cq X be a Laurent polynomial in Q[X,X
By the support of f we understand the subset of Zn defined by the elements
q ∈ Zn for which cq 6= 0 holds. The Newton polytope of f is the convex hull
of the support of f in Rn .
A sparse polynomial system with supports ∆1 , . . . , ∆n ⊂ (Z≥0 )n is defined
by polynomials
X
fi (X) :=
ai,q X q (1 ≤ i ≤ n),
q∈∆i
with ai,q ∈ C \ {0} for each q ∈ ∆i and 1 ≤ i ≤ n.
For a finite subset ∆ of Zn , we denote by Q := Conv(∆) its convex hull
124
in Rn . The usual Euclidean volume of a polytope Q in Rn will be denoted
by VolRn (Q).
Let Q1 , . . . , Qn be polytopes in Rn . For λ1 , . . . , λn ∈ R≥0 , we use the notation λ1 Q1 + · · · + λn Qn to refer to the Minkowski sum λ1 Q1 + · · · + λn Qn :=
{x ∈ Rn : x = λ1 x1 + · · · + λn xn with x1 ∈ Q1 , . . . , xn ∈ Qn }. Consider the
real-valued function (λ1 , . . . , λn ) 7→ VolRn (λ1 Q1 +· · ·+λn Qn ). This is a homogeneous polynomial function of degree n in the λi (see, e.g., [CLO98, Chapter
7, Proposition §4.4.9]). The mixed volume M(Q1 , . . . , Qn ) of Q1 , . . . , Qn is defined as the coefficient of the monomial λ1 · · · λn in VolRn (λ1 Q1 +· · ·+λn Qn ).
For i = 1, . . . , n, let ∆i be a finite subset of Zn≥0 and let Qi := Conv(∆i )
denote the corresponding polytope. Let f1 , . . . , fn be a sparse polynomial
system with respect to ∆1 , . . . , ∆n . The BKK Theorem ([Ber75], [Kus76],
[Kho78]) asserts that the system f1 = 0, . . . , fn = 0 has at most M(Q1 , . . . , Qn )
isolated common solutions in the n-dimensional torus (C∗ )n , with equality
for generic choices of the coefficients of f1 , . . . , fn . Furthermore, if the condition 0 ∈ Qi holds for 1 ≤ i ≤ n, then M(Q1 , . . . , Qn ) bounds the number
of solutions in Cn (see [LW96]).
Example. Let ∆1 := {(0, 0), (2, 0), (0, 2), (2, 2)} and ∆2 := {(0, 0), (1, 2), (2, 1)}
in Z2 . A sparse polynomial system with supports ∆1 , ∆2 is a system defined
by polynomials of the following type:
(
f1 = a(0,0) + a(2,0) X12 + a(0,2) X22 + a(2,2) X12 X22 ,
(5.1)
f2 = b(0,0) + b(1,2) X1 X22 + b(2,1) X12 X2 ,
with aq , bq ∈ C \ {0}.
Let Q1 := Conv(∆1 ) and Q2 := Conv(∆2 ). Then M(Q1 , Q2 ) = VolR2 (Q1 +
Q2 ) − VolR2 (Q1 ) − VolR2 (Q2 ) = 8.
The pictures of Q1 , Q2 and Q1 + Q2 are respectively:
5.1.1.
125
Mixed subdivisions
Assume that the union of the sets ∆1 , . . . , ∆n affinely generates Zn , and
consider the partition of ∆1 , . . . , ∆n defined by the relation ∆i ∼ ∆j if and
only if ∆i = ∆j . Let s ∈ N denote the number of classes in this partition,
and let A(1) , . . . , A(s) ⊂ Zn denote a member in each class. Write A :=
(A(1) , . . . , A(s) ). For ℓ = 1, . . . , s, let kℓ := #{i : ∆i = A(ℓ) }. Without loss
of generality, we will assume that ∆1 = · · · = ∆k1 = A(1) , ∆k1 +1 = · · · =
∆k1 +k2 = A(2) and so on.
A cell of A is a tuple C = (C (1) , . . . , C (s) ) with C (ℓ) 6= ∅ and C (ℓ) ⊂ A(ℓ)
for 1 ≤ ℓ ≤ s. We define
type(C) := (dim(Conv(C (1) )), . . . , dim(Conv(C (s) ))),
Conv(C) := Conv(C (1) + · · · + C (s) ),
#(C) := #(C (1) ) + · · · + #(C (s) ),
VolRn (C) := VolRn (Conv(C)).
A face of a cell C is a cell C = (C (1) , . . . , C (s) ) of C with C (ℓ) ⊂ C (ℓ) for
1 ≤ ℓ ≤ s such that there exists a linear functional γ : Rn → R that takes
its minimum over C (ℓ) at C (ℓ) for 1 ≤ ℓ ≤ s. One such functional γ is called
an inner normal of C.
A mixed subdivision of A is a collection of cells C = {C1 , . . . , Cm } of A
satisfying conditions (1)–(4) below.
1. dim(Conv(Cj )) = n for 1 ≤ j ≤ m,
2. the intersection Conv(Ci ) ∩ Conv(Cj ) ⊂ Rn is either the empty set or
a face of both Conv(Ci ) and Conv(Cj ) for 1 ≤ i < j ≤ m,
S
3. m
j=1 Conv(Cj ) = Conv(A),
4.
Ps
ℓ=1
(ℓ)
dim(Conv(Cj )) = n for 1 ≤ j ≤ m.
If C also satisfies the condition
5. #(Cj ) = n + s for 1 ≤ j ≤ m,
126
we say that C is a fine-mixed subdivision of A. Observe that, as a consequence
(1)
(s)
of conditions (4) and (5), for each cell Cj = (Cj , . . . , Cj ) in a fine-mixed
(ℓ)
(ℓ)
subdivision the identity dim(Conv(Cj )) = #Cj − 1 holds for 1 ≤ ℓ ≤ s.
In the sequel, we are going to consider only cells of type (k1 , . . . , ks ) in a
fine-mixed subdivision.
We point out that a mixed subdivision C of A enables us to compute the
mixed volume of the family Q1 = Conv(∆1 ), . . . , Qn = Conv(∆n ) by means
of the following identity (see [HS95, Theorem 2.4.]):
X
M(Q1 , . . . , Qn ) =
k1 ! . . . ks ! · VolRn (Ci ).
(5.2)
Ci ∈C
type(Ci )=(k1 ,...,ks )
A fine-mixed subdivision of A can be obtained by means of a lifting
process as explained in what follows. For 1 ≤ ℓ ≤ s, let ωℓ : A(ℓ) → R be
an arbitrary function. The tuple ω := (ω1 , . . . , ωs ) is called a lifting function
for A. Once a lifting function ω is fixed, the graph of any subset C (ℓ) of
b(ℓ) := {(q, ωℓ (q)) ∈ Rn+1 : q ∈ C (ℓ) }. Then, for a
A(ℓ) will be denoted by C
sufficiently generic lifting function ω, the set of cells C of A satisfying the
following conditions.
b(1) + · · · + C
b(s) )) = n,
i. dim(Conv(C
b(1) , . . . , C
b(s) ) is a face of (Ab(1) , . . . , Ab(s) ) whose inner normal has posii. (C
itive last coordinate,
is a fine-mixed subdivision of A (see [HS95, Section 2]).
Example. We continue with the example introduced at the end of the previous
subsection. Here A := (A(1) , A(2) ), where A(1) := ∆1 and A(2) := ∆2 .
Following [HS95, Example 2.7], the lifting function ω = (ω1 , ω2 ) defined
by
ω1 (q) :=
(
0 for q = (0, 0)
1 for q ∈ A(1) \ {(0, 0)}
and ω2 (q) := 0 for every q ∈ A(2) ,
(5.3)
induces a fine-mixed subdivision of A. More precisely, such a fine–mixed
subdivision consists of the set of cells satisfying conditions (i) and (ii) above,
127
which are listed below together with the inner normals of the faces they come
from:
C1 := {{(0, 0), (0, 2)}, {(0, 0), (1, 2)}}, γ (1) := (2, −1, 2).
C2 := {{(0, 0), (2, 0)}, {(0, 0), (2, 1)}}, γ (2) := (−1, 2, 2).
C3 := {{(0, 0), (2, 2)}, {(1, 2), (2, 1)}}, γ (3) := (−1, −1, 4).
C4 := {{(0, 0)}, {(0, 0), (1, 2), (2, 1)}}, γ (4) := (0, 0, 1).
C5 := {{(0, 0), (0, 2), (2, 2)}, {(1, 2)}}, γ (5) := (0, −1, 2).
C6 := {{(0, 0), (2, 0), (2, 2)}, {(2, 1)}}, γ (6) := (−1, 0, 2).
The pictures below show the lower envelope of Ab(1) + Ab(2) ⊂ R3 and its
projection to R2 respectively.
Note that the cells of type (k1 , k2 ) = (1, 1) are C1 , C2 and C3 .
The following result (cf. [HS95, Section 2]) states a generic condition for
a lifting function to induce a fine-mixed subdivision:
Lemma 5.2. The lifting process associated to a lifting function ω yields
a fine-mixed subdivision
holds: for every
Ps of A if the following condition
(1)
r1 , . . . , rs ∈ Z≥0 with ℓ=1 rℓ > n and every cell (C , . . . , C (s) ) with C (ℓ) :=
128
{qℓ,0 , . . . , qℓ,rℓ } ⊂ A(ℓ) (1 ≤ ℓ ≤ s), if



q1,1 − q1,0
..



.






 q1,r1 − q1,0 




···



b
V (C) := 
and
V
(
C)
:=


···







 qs,1 − qs,0 



.



..
qs,rs − qs,0
q1,1 − q1,0
..
.
q1,r1 − q1,0
···
···
qs,1 − qs,0
..
.
qs,rs − qs,0

ω1 (q1,1 ) − ω1 (q1,0 )
..

.


ω1 (q1,r1 ) − ω1 (q1,0 ) 

···

,
···


ωs (qs,1 ) − ωs (qs,0 ) 

..

.
ωs (qs,rs ) − ωs (qs,0 )
b = n + 1.
then rank(V (C)) = n implies rank(V (C))
b = n + 1 can be restated as the nonNote that the condition rank(V (C))
b Since rank(V (C)) = n,
vanishing of the maximal minors of the matrix V (C).
these maximal minors are nonzero linear forms in the unknown values ωℓ (qℓ,j )
of the lifting function. Thus, if Nℓ = #A(ℓ) for every 1 ≤ ℓ ≤ s, a sufficiently
generic lifting function can be obtained by randomly choosing the values
ωℓ (qℓ,j ) of ω at the points of A(ℓ) from the set {1, 2, . . . , ρ2N1 +···+Ns }, with
probability of success at least 1 − 1/ρ for ρ ∈ N.
In this chapter, we shall assume given a sufficiently generic lifting function
and the induced fine-mixed subdivision of A.
5.1.2.
Degree estimates in the sparse setting
Suppose that we are given a curve V ⊂ An+1 (C) defined by polynomials
f1 , . . . , fn ∈ Q[X, T ]. Assume that for each irreducible component C of V ,
the identity I(C) ∩ Q[T ] = {0} holds. Let u be a nonzero linear form of Q[X]
and πu : V → A2 the morphism defined by πu (x, t) := (t, u(x)).
Our assumptions on V imply that the Zariski closure πu (V ) of the image
of V under πu is an hypersurface of A2 defined over Q, and hence, there exists
a unique (up to scaling by nonzero elements of Q) polynomial Mu ∈ Q[T, Y ]
of minimal degree defining πu (V ). Let mu ∈ Q(T )[Y ] denote the (unique)
monic multiple of Mu with degY (mu ) = degY (Mu ); so mu is the minimal
polynomial of u in V .
129
As in the previous chapters, the complexity of the algorithms for solving
a system f1 = 0, . . . , fn = 0 defining such a curve can be expressed mainly
by means of the degree and height of the projection πu : V → A1 . The degree
of πu is equal to the degree deg mu = degY Mu of the minimal polynomial of
a generic linear form u ∈ Q[X1 , . . . , Xn ] and the height of πu is equal to the
partial degree degT Mu (see Section 2.4).
In the sparse setting, we can estimate degY Mu and degT Mu in terms of
combinatorial quantities (namely, mixed volumes) associated to the polynomial system under consideration (see also [PS08]).
Lemma 5.3. Let assumptions and notations be as above. For 1 ≤ i ≤ n,
let Qi ⊂ Rn be the Newton polytope of fi , considering fi as an element of
b1 , . . . , Q
bn ⊂ Rn+1 be the Newton polytopes of f1 , . . . , fn , conQ(T )[X]. Let Q
sidering f1 , . . . , fn as elements of Q[X, T ], and let ∆ ⊂ Rn+1 be the standard
n-dimensional simplex in the hyperplane {T = 0}, i.e., the Newton polytope
bi for every 1 ≤ i ≤ n.
of a generic linear form u ∈ Q[X]. Assume that 0 ∈ Q
Then the following estimates hold:
b1 , . . . , Q
bn ).
degY Mu ≤ M(Q1 , . . . , Qn ), degT Mu ≤ M(∆, Q
(5.4)
bi ⊂ Qi × [0, ci ] for
Furthermore, if there exist c1 , . . . , cn ∈ R≥0 such that Q
1 ≤ i ≤ n, then the following inequality holds:
degT Mu ≤
n
X
ci M(∆, Q1 , . . . , Qi−1 , Qi+1 , . . . , Qn ).
(5.5)
i=1
Proof. The upper bound degY Mu ≤ M(Q1 , . . . , Qn ) follows straightforwardly from the BKK bound and the affine root count in [LW96].
In order to obtain an upper bound for degT Mu , we observe that substituting a generic value y ∈ Q for Y we have degT Mu (T,Y ) = degT Mu (T, y) =
#{t ∈ C; Mu (t, y) = 0}. Moreover, it follows that Mu (t, y) = 0 if and
only if there exists a point x ∈ An with y = u(x) and (x, t) ∈ V . Therefore, it suffices to estimate the number of points (x, t) ∈ An+1 satisfying
u(x) − y = 0, f1 (x, t) = 0, . . . , fn (x, t) = 0. Being u a generic linear form, the
system
u(X) − y = 0, f1 (X, T ) = 0, . . . , fn (X, T ) = 0
(5.6)
has finitely many common zeros in An+1 . Combining the BKK bound with
b1 , . . . , Q
bn )
the affine root count of [LW96] we see that there are at most M(∆, Q
130
b1 , . . . , Q
bn ) holds,
solutions of (5.6). We conclude that degT Mu ≤ M(∆, Q
showing thus (5.4).
In order to prove (5.5), we make use of basic properties of the mixed
bi ⊂ Qi × [0, ci ] holds for
volume (see, for instance, [Ewa96, Ch. IV]). Since Q
1 ≤ i ≤ n, by the monotonicity of the mixed volume we have
b1 , . . . , Q
bn ) ≤ M(∆, Q1 × [0, c1 ], . . . , Qn × [0, cn ]).
M(∆, Q
Note that Qi ×[0, ci ] = Si,0 +Si,1 , where Si,0 = Qi ×{0} and Si,1 = {0}×[0, ci ]
for i = 1, . . . , n. Hence, by multilinearity,
X
M(∆, S1,j1 , . . . , Sn,jn ). (5.7)
M(∆, Q1 × [0, c1 ], . . . , Qn × [0, cn ]) =
(j1 ,...,jn )∈{0,1}n
If the vector (j1 , . . . , jn ) has at least two nonzero coordinates, then two of
the sets S1,j1 , . . . , Sn,jn are parallel line segments; therefore, M(∆, S1,j1 , . . . ,
Sn,jn ) = 0. On the other hand, if ji is the only nonzero coordinate, the
corresponding term in the sum of the right-hand side of (5.7) is
M(∆, Q1 × {0}, . . . , Qi−1 × {0}, {0} × [0, ci ], Qi+1 × {0}, . . . , Qn × {0})
= ci M(∆, Q1 , . . . , Qi−1 , Qi+1 , . . . , Qn ).
Finally, for (j1 . . . , jn ) = (0, . . . , 0) we have M(∆, Q1 ×{0}, . . . , Qn ×{0}) = 0
since all the polytopes are included in an n-dimensional subspace.
We conclude that the right-hand side of (5.7) equals the right-hand side
of (5.5). This finishes the proof of the lemma.
Example. For the system
(
a(0,0) + a(2,0) X12 T + a(0,2) X22 T + a(2,2) X12 X22 T = 0,
b(0,0) + b(1,2) X1 X22 + b(2,1) X12 X2 = 0,
we have:
Q1 = Conv({(0, 0), (0, 2), (2, 0), (2, 2)}),
Q2 = Conv({(0, 0), (1, 2), (2, 1)}),
(5.8)
5.2. SOLUTION OF A GENERIC SPARSE SYSTEM
131
b1 = Conv({(0, 0, 0), (0, 2, 1), (2, 0, 1), (2, 2, 1)}),
Q
b2 = Conv({(0, 0, 0), (1, 2, 0), (2, 1, 0)}).
Q
Therefore, the following upper bounds for the degree of the polynomial Mu
hold for any separating linear form u:
degY Mu ≤ M(Q1 , Q2 ) = 8 =: D,
b1 , Q
b2 ) = 3 =: E,
degT Mu ≤ M(∆, Q
(5.9)
(5.10)
where ∆ := Conv({(0, 0, 1), (1, 0, 0), (0, 1, 0)}).
5.2.
Solution of a generic sparse system
Let ∆1 , . . . , ∆n be fixed finite subsets of Zn≥0 with 0 ∈ ∆i for 1 ≤ i ≤ n
and let D := M(Q1 , . . . , Qn ) denote the mixed volume of the polytopes Q1 :=
Conv(∆1 ),P
. . . , Qn := Conv(∆n ). Assume that D > 0 holds or, equivalently,
that dim
i∈I Qi ≥ |I| for every non-empty subset I ⊂ {1, . . . , n} (see, for
instance, [Oka97, Chapter IV, Proposition 2.3]).
Let f1 , . . . , fn ∈ Q[X] be polynomials defining a sparse system with
respect to ∆1 , . . . , ∆n and let d1 , . . . , dn be their total degrees. Let d :=
max{d1 , . . . , dn }. Suppose that f1 , . . . , fn define a zero-dimensional variety
V in An . Group equal supports into s ≤ n distinct supports and define
representatives as A(1) , . . . , A(s) as we did in the previous section. Write
A := (A(1) , . . . , A(s) ) and denote by kℓ the number of polynomials fi with
support A(ℓ) for 1 ≤ ℓ ≤ s.
From now on we assume that we are given a sufficiently generic lifting
function ω := (ω1 , . . . , ωs ) and the fine-mixed subdivision of A induced by
ω. We assume further that, for every 1 ≤ ℓ ≤ s, the function ωℓ : A(ℓ) → Z
takes only nonnegative values and ωℓ (0, . . . , 0) = 0.
We introduce auxiliary generic polynomials g1 , . . . , gn with the same supports ∆1 , . . . , ∆n (satisfying a geometric condition to be made explicit in
Section 5.2.1) and consider the perturbed polynomial system defined by
h1 := f1 + g1 , . . . , hn := fn + gn . We observe that if the coefficients of the
polynomials f1 , . . . , fn satisfy this condition then our symbolic adaptation of
Huber-Sturmfels method can be applied directly to f1 , . . . , fn . Otherwise, we
132
first solve the system h1 = 0, . . . , hn = 0 using the method to be discussed below and then recover the solutions to the input system f1 = 0, . . . , fn = 0 by
considering the standard homotopy f1 +(1−T )g1 = 0, . . . , fn +(1−T )gn = 0.
5.2.1.
The polyhedral deformation
This section is devoted to introducing the polyhedral deformation of Huber and Sturmfels.
LetPus maintain notions and notations from the previous section. Let
hi := q∈∆i ci,q X q for 1 ≤ i ≤ n be polynomials in Q[X], where ∆1 , . . . , ∆n
are the supports introduced in Section 5.2 and let V1 denote the set of their
common zeros in An . For i = 1, . . . , n, let ℓi be the (unique) integer with
∆i = A(ℓi ) , and let ω
ei := ωℓi be the lifting function associated to the support
∆i . In order to simplify notations, the n-tuple ω
e := (e
ω1 , . . . , ω
en ) will be
denoted simply by ω = (ω1 , . . . , ωn ).
b(ℓ) := {(q, ωℓ (q)) ∈ Rn+1 : q ∈ C (ℓ) }
As done before, we denote by C
the graph of any subset C (ℓ) of A(ℓ) for 1 ≤ ℓ ≤ s, and extend this notation
correspondingly. Let T be a new indeterminate. We consider a “deformation”
of the polynomials h1 , . . . , hn into polynomials b
h1 , . . . , b
hn ∈ Q[X, T ] defined
in the following way:
X
b
hi (X, T ) :=
ci,q X q T ωi (q) for 1 ≤ i ≤ n.
(5.11)
q∈∆i
Let I denote the ideal of Q[X, T ] generated by b
h1 , . . . , b
hn and let J denote the
b
b
Jacobian determinant of h1 , . . . , hn with respect to the variables X1 , . . . , Xn .
We set
Vb := V (I : J ∞ ) ⊂ An+1 .
(5.12)
We shall show that, under a generic choice of the coefficients of h1 , . . . , hn , the
system defined by the polynomials in (5.11) constitutes a deformation of the
input system h1 = 0, . . . , hn = 0, in the sense that the morphism π : Vb → A1
defined by π(x, t) := t is a dominant map with π −1 (1) = V1 × {1}.
We shall further exhibit degree estimates on the genericity condition underlying such choice of coefficients. These estimates will allow us to obtain
suitable polynomials h1 , . . . , hn by randomly choosing their coefficients in an
appropriate finite subset of Z.
133
According to [HS95, Section 3], the solutions over an algebraic closure
Q(T ) of Q(T ) to the system defined by the polynomials (5.11) are algebraic
functions of the parameter T which can be represented as Puiseux series of
the form
γ1
γn
x(T ) := (x1,0 T γn+1 + higher-order terms, . . . , xn,0 T γn+1 + higher-order terms),
(5.13)
n+1
where γ := (γ1 , . . . , γn , γn+1 ) ∈ Z
is an inner normal with positive last
b = (C
b(1) , . . . , C
b(s) ) of Ab of type
coordinate γn+1 > 0 of a (lower) facet C
(k1 , . . . , ks ), and x0 := (x1,0 , . . . , xn,0 ) ∈ (C∗ )n is a solution to the polynomial
system defined by
X
(0)
hi,γ :=
ci,q X q
(1 ≤ i ≤ n),
(5.14)
q∈C (ℓi )
where, as defined before, ℓi is the integer with 1 ≤ ℓi ≤ s such that ∆i = A(ℓi ) .
For a generic choice of the coefficients of the polynomials h1 , . . . , hn there are
k1 ! · · · ks ! · Vol(C) distinct solutions x0 ∈ (C∗ )n to the system defined by the
polynomials (5.14) and hence, there are k1 ! · · · ks ! · Vol(C) distinct Puiseux
series x(T ) as in (5.13).
We shall “lift” each of these solutions x0 to a solution of the form (5.13)
to the system defined by (5.11). Explicitly, on input x0 , we shall compute
the Puiseux series expansion of the corresponding solution (5.13) truncated
up to a suitable order.
Let
(0)
V0,γ := {x ∈ (C∗ )n : h1,γ (x) = 0, . . . , h(0)
n,γ (x) = 0}.
(5.15)
A particular feature of the polynomials (5.14) which makes the associated
equation system “easy to solve” is that the vector of their supports is (C (1) )k1 ×
· · ·×(C (s) )ks , where (C (1) , . . . , C (s) ) is a cell of type (k1 , . . . , ks ) in a fine-mixed
subdivision of A. Therefore, for every 1 ≤ ℓ ≤ s, the set C (ℓ) consists of kℓ +1
points and hence, the (Laurent) polynomials in (5.14) are linear combinations
of n + 1 distinct monomials in n variables (up to monomial multiplication so
that each polynomial has a nonzero constant term).
Denote Γ ⊂ Zn+1 the set of all primitive integer vectors of the form
γ := (γ1 , . . . ,γn ,γn+1 ) ∈ Zn+1 with γn+1 > 0 for which there is a cell C =
(C (1) , . . . , C (s) ) of type (k1 , . . . , ks ) of the subdivision of A induced by ω such
b has inner normal γ.
that C
134
Fix a cell C = (C (1) , . . . , C (s) ) of type (k1 , . . . , ks ) of the subdivision of A
induced by ω associated with a primitive inner normal γ ∈ Γ with positive
last coordinate. In order to lift the points of the variety V0,γ of (5.15) to a
solution of the system defined by the polynomials in (5.11), we will work with
a family of auxiliary polynomials h1,γ , . . . , hn,γ ∈ Q[X, T ] which we define as
follows:
hi (T γ1 X1 , . . . , T γn Xn , T γn+1 ) (1 ≤ i ≤ n)
hi,γ (X, T ) := T −mi b
(5.16)
where mi ∈ Z is the lowest power of T appearing in b
hi (T γ1 X1 , . . . , T γn Xn , T γn+1 )
for every 1 ≤ i ≤ n. Note that the polynomials obtained by substituting
T = 0 into h1,γ , . . . , hn,γ are precisely those introduced in (5.14). Now we
illustrate the objects introduced in this subsection with a particular sparse
polynomial system with the same structure as the generic system (5.1).
Example. Consider the polynomials h1 , h2 ∈ Q[X1 , X2 ] defined as:
(
h1 := 1 − X12 − X22 − X12 X22 ,
h2 := 1 + X12 X2 + X1 X22 .
(5.17)
Observe that the polynomials above are a specialization of the generic polynomials introduced in (5.1).
We deform the polynomials h1 , h2 using the lifting function ω defined in
(5.3), obtaining thus the following polynomials:
(
b
h1 := 1 − X12 T − X22 T − X12 X22 T,
(5.18)
b
h2 := 1 + X12 X2 + X1 X22 .
These polynomials b
h1 , b
h2 define the curve
Vb := V ((b
h1 , b
h2 ) : J ∞ ) = V (b
h1 , b
h2 ),
(5.19)
where J is the Jacobian determinant of b
h1 and b
h2 with respect to the variables
X 1 , X2 .
According to the remark at the end of the example of Subsection 5.1.1, the
cells of type (1, 1) in the fine-mixed subdivision of the support sets induced
by ω, and the corresponding inner normals are:
135
C1 := {{(0, 0), (0, 2)}, {(0, 0), (1, 2)}}, γ (1) := (2, −1, 2).
C2 := {{(0, 0), (2, 0)}, {(0, 0), (2, 1)}}, γ (2) := (−1, 2, 2).
C3 := {{(0, 0), (2, 2)}, {(1, 2), (2, 1)}}, γ (3) := (−1, −1, 4).
(0)
Therefore, the polynomial systems defined by the polynomials hi,γ of (5.14)
and their solution sets V0,γ are:
( (0)
h1,γ (1) = 1 − X22 ,
V0,γ (1) = {(−1, 1), (−1, −1)},
(5.20)
(0)
h2,γ (1) = 1 + X1 X22 ,
(
(0)
h1,γ (2) = 1 − X12 ,
(0)
h2,γ (2) = 1 + X12 X2 ,
( (0)
h1,γ (3) = 1 − X12 X22 ,
(0)
V0,γ (2) = {(1, −1), (−1, −1)},
h2,γ (3) = X12 X2 + X1 X22 ,
(5.21)
V0,γ (3) = {(1, −1), (−1, 1), (i, −i), (−i, i)}.
(5.22)
Finally, the polynomials hi,γ defined in (5.16) are:
(
h1,γ (1) = 1 − X12 T 6 − X22 − X12 X22 T 4 ,
h2,γ (1) = 1 + X12 X2 T 3 + X1 X22 ,
(
(
h1,γ (2) = 1 − X12 − X22 T 6 − X12 X22 T 4 ,
h2,γ (2) = 1 + X12 X2 + X1 X22 T 3 ,
(5.23)
(5.24)
h1,γ (3) = 1 − X12 T 2 − X22 T 2 − X12 X22 ,
h2,γ (3) = T 3 + X12 X2 + X1 X22 .
(5.25)
5.2.2.
On the genericity of the initial system
Here we discuss the genericity conditions underlying the choice of the
polynomials g1 , . . . , gn that enable us to apply the polyhedral deformation
defined by the lifting form ω to the system h1 := f1 + g1 = 0, . . . , hn :=
fn + gn = 0.
136
The first condition we require is that the set of common zeros of the perturbed polynomials h1 , . . . , hn is a zero-dimensional variety with the maximum number of points for a sparse system with the given structure. More
precisely, we require that the following condition holds.
(H1) The set V1 := {x ∈ An : h1 (x) = 0, . . . , hn (x) = 0} is a zero-dimensional
variety with D := M(Q1 , . . . , Qn ) distinct points.
In addition, we need that the system (5.14) giving the initial points to
our first deformation for every γ ∈ Γ has as many roots as possible, namely
the mixed volume of their support vectors.
For each cell C = (C (1) , . . . , C (s) ) of type (k1 , . . . , ks ) of the induced finemixed subdivision, set an order on the n + 1 points appearing in any of the
sets C (ℓ) , after a suitable translation so that 0 ∈ C (ℓ) for every 1 ≤ ℓ ≤ s.
Assume that 0 ∈ Zn is the last point according to this order. Denote γ ∈ Zn+1
the primitive inner normal of C with positive last coordinate. Consider the
(0)
n × (n + 1) matrix whose ith row is the coefficient vector of hi,γ in the
prescribed monomial order and set Mγ ∈ Qn×n and Bγ ∈ Qn×1 for the
submatrices consisting of the first n columns (coefficients of non-constant
monomials) and the last column (constant coefficients) respectively. Then,
the coefficients of g1 , . . . , gn are to be chosen so that the following condition
holds.
(H2) For every γ ∈ Γ, the (n × n)-matrix Mγ is nonsingular and all the
entries of (Mγ )−1 Bγ are nonzero.
Our next results assert that the above conditions can be met with good
probability by randomly choosing the coefficients of g1 , . . . , gn in a certain
set S ⊂ Z. We observe that our estimate on the size of S is not intended to
be accurate, but to show that the growth of the size of the integers involved
in the subsequent computations is not likely to create complexity problems.
Let {Ωi,q : 1 ≤ i ≤ n, q ∈ ∆i } be a set of new indeterminates over Q. For
1 ≤ i ≤ n, write Ωi := (Ωi,q : q ∈ ∆i ) and let Hi ∈ Q[Ωi , X] be the generic
polynomial
X
Hi (Ωi , X) :=
Ωi,q X q
(5.26)
q∈∆i
137
with support ∆i and Ni := #∆i coefficients. Let Ω := (Ω1 , . . . , Ωn ) and let
N := N1 + · · · + Nn be the total number of indeterminate coefficients.
We start the analysis of the required generic conditions with the following
quantitative version of Bernstein’s result on the genericity of zero-dimensional
sparse systems (see [Ber75, Theorem B], [HS95, Theorem 6.1]).
Lemma 5.4. There exists a nonzero polynomial P (0) ∈ Q[Ω] with deg P (0) ≤
3n2n+1 d2n−1 such that for any c ∈ QN with P (0) (c) 6= 0, the system H1 (c1 , X) =
0, . . . , Hn (cn , X) = 0 has D solutions in An , counting multiplicities.
Proof. Due to [HS95, Theorem 6.1] combined with [LW96], the system
H1 (c1 , X) = 0, . . . , Hn (cn , X) = 0 has D solutions in An counting multiplicities if and only if for every facet inner normal µ ∈ Zn of Q1 + · · · + Qn ,
the sparse resultant Res∆µ1 ,...,∆µn does not vanish at c := (c1 , . . . , cn ). Here
∆µi denotes the set of points of ∆i where the linear functional induced by µ
attains its minimum for 1 ≤ i ≤ n.
Q
Therefore, the polynomial P (0) := µ Res∆µ1 ,...,∆µn ∈ Q[Ω], where the product ranges over all primitive inner normals µ ∈ Zn to facets of Q1 + · · · + Qn ,
satisfies the required condition.
In order to estimate the degree of P (0) , we observe that for every facet
inner normal µ ∈ Zn the following upper bound holds:
deg(Res
µ
∆µ
1 ,...,∆n
)≤
n
X
Mn−1 (∆µ1 , . . . , ∆µi−1 , ∆µi+1 , . . . , ∆µn ) ≤ ndn−1 ,
i=1
where d := max{d1 , . . . , dn }. On the other hand, it is not difficult to see that
the number of facets of an n-dimensional integer convex polytope P ⊂ Rn
which has an integer point in its interior is bounded by n! VolRn (P ). Now,
taking P := (n + 1)Q, we obtain an integer polytope with the same number
of facets as Q having an integer interior point. Then, the number of facets
of Q is bounded by
n!VolRn (P ) = n! VolRn ((n + 1)Q)
= (n + 1)n n! VolRn (Q)
≤ (n + 1)n (nd)n ,
138
since Q is included in the n-dimensional simplex of size nd. This proves the
upper bound for the degree P (0) of the statement of the lemma.
The next lemma is concerned with the genericity of a smooth sparse
system.
Lemma 5.5. With the same notations as in Lemma 5.4 and before, there exists a nonzero polynomial P (1) ∈ Q[Ω] of degree at most 4n2n+1 d2n−1 such that
for any c ∈ QN with P (1) (c) 6= 0, the system H1 (c1 , X) = 0, . . . , Hn (cn , X) =
0 has exactly D distinct solutions in An .
Proof. Consider the incidence variety associated to (∆1 , . . . , ∆n )-sparse systems, namely
X
W := {(x, c) ∈ (C∗ )n × (AN1 × · · · × ANn ) :
ci,q xq = 0 for 1 ≤ i ≤ n}.
q∈∆i
As in [PS93, Proposition 2.3], it follows that W is a Q-irreducible variety. Let
πΩ : W → AN1 × · · · × ANn be the canonical projection, which is a dominant
map.
By [Oka97, Chapter V, Corollary (3.2.1)], there is a nonempty Zariski
open set U (∆1 , . . . , ∆n ) ⊂ AN1 × · · · × ANn of coefficients c = (c1 , . . . , cn ) for
which the polynomials H1 (c1 , X), . . . , Hn (cn , X) have supports ∆1 , . . . , ∆n
respectively and the set of their common zeros in (C∗ )n is a non-degenerate
complete intersection variety. Then, the Jacobian JH := det(∂Hi /∂Xj )1≤i,j≤n
does not vanish at any point of πΩ−1 (c) for every c ∈ U(∆1 , . . . , ∆n ).
Let Q(Ω) ֒→ Q(W ) be the finite field extension induced by the dominant
map πΩ . By the preceding paragraph we have that the rational function defined by JH in Q(W ) is nonzero. Therefore, its primitive minimal polynomial
MJ ∈ Q[Ω, Y ] is well defined and satisfies the degree estimates
degΩ MJ ≤ deg W · deg JH ≤
n
Y
i=1
(di + 1) ·
n
X
di ≤ 2n dn+1 n
i=1
(see [SS96a], [Sch03]).
(0)
Let P (1) := P (0) MJ , where P (0) is the polynomial given by Lemma 5.4
(0)
and MJ denotes the constant term of the expansion of MJ in powers of
139
Y . We claim that P (1) satisfies the requirements of the statement of the
lemma. Indeed, let c ∈ QN satisfy P (1) (c) 6= 0. Then P (0) (c) 6= 0 holds
and hence, Lemma 5.4 implies that H1 (c, X) = 0, . . . , Hn (c, X) = 0 is a
(0)
zero-dimensional system. Furthermore, MJ (c) is a nonzero multiple of the
Q
(0)
product x∈π−1 (c) JH (c, x). Thus, the non-vanishing of MJ (c) shows that
Ω
all the points of πΩ−1 (c) are smooth and therefore, from e.g. [Oka97, IV,
Theorem 2.2], it follows that πΩ−1 (c) consists of exactly D simple points in
(C∗ )n . Moreover, combining the assumption that 0 ∈ ∆i for 1 ≤ i ≤ n with
[LW96, Theorem 2.4], we deduce that πΩ−1 (c) consists of D simple points in
(0)
An . The estimate deg MJ ≤ degΩ MJ ≤ 2n dn+1 n ≤ n2(n+1) d2n−1 implies the
statement of the lemma.
Finally, we exhibit a generic condition on the coefficients h1 , . . . , hn which
implies that assumption (H2) holds.
Lemma 5.6. With the previous assumptions and notations, there exists a
nonzero polynomial P (2) ∈ Q[Ω] with deg P (2) ≤ n(n + 1)#Γ such that for
every c := (c1 , . . . , cn ) ∈ QN with P (2) (c) 6= 0, the polynomials hi := Hi (ci , X)
(1 ≤ i ≤ n) satisfy condition (H2).
b
Proof. Fix a primitive integer inner normal γ ∈ Γ to a lower facet of A.
n×n
n×1
Let Mγ ∈ Q[Ω]
and Bγ ∈ Q[Ω]
be the matrices constructed from the
generic polynomials H1 , . . . , Hn ∈ Q[Ω][X] as explained in the paragraph
preceding condition (H2). Let D0,γ ∈ Q[Ω] be the (nonzero) determinant
of Mγ , and for every 1 ≤ j ≤ n, let Dj,γ be the determinant of the
Qn matrix
obtained from Mγ by replacing its jth column with Bγ . Set Pγ := j=0 Dj,γ .
Q
Finally, take P (2) := γ∈Γ Pγ . By Cramer’s rule, whenever P (2) (c) 6= 0, we
have that the system h1 , . . . , hn with coefficient vector c = (c1 , . . . , cn ) meets
condition (H2).
The degree estimate for P (2) follows from the fact that deg Pγ ≤ n(n +
1) holds for every γ ∈ Γ, since each of the entries of the matrices whose
determinants are involved has degree 1 in the variables Ω.
Now, we are ready to state a generic condition on the coefficients of
h1 , . . . , hn which implies that (H1) and (H2) hold.
Proposition 5.7. Under the previous assumptions and notations, there exists a nonzero polynomial P ∈ Q[Ω] with deg P ≤ 4n2n+1 d2n−1 + n(n + 1)D
140
such that for every c ∈ QN with P (c) 6= 0, the polynomials hi := Hi (ci , X)
(1 ≤ i ≤ n) satisfy conditions (H1) and (H2).
Proof. Set P := P (1) P (2) , where P (1) is the polynomial of the statement of
Lemma 5.5 and P (2) is the one defined in the statement of Lemma 5.6. The
result follows from Lemmas 5.5 and 5.6, and the upper bound #Γ ≤ D for
the cardinality of the set of the distinct inner normal vectors considered (one
for each cell of type (k1 , . . . , ks ) in the given fine-mixed subdivision).
5.2.3.
Outline of the algorithm
Now we have all the tools necessary to give an outline of our algorithm for
the computation of a geometric solution of the (sufficiently generic) sparse
system h1 = 0, . . . , hn = 0.
With notations as in the previous subsections, we assume that a finemixed subdivision of A induced by a lifting function ω is given. This means
that we are given the set Γ of inner normals of the lower facets of the convex
b together with the corresponding cells of the convex hull of A. In
hull of A,
addition, we suppose that our input polynomials h1 , . . . , hn ∈ Q[X] satisfy
conditions (H1) and (H2) and denote by V1 ⊂ An the affine variety defined
by h1 , . . . , hn .
First, we choose a generic linear form u ∈ Q[X] such that:
u separates the points of the zero-dimensional varieties V1 and V0,γ for
every γ ∈ Γ. This condition is represented by the nonvanishing of a
certain nonconstant polynomial of degree at most 4D2 .
An algorithm for the computation of the minimal polynomial of u in
V0,γ described below can be extended to a computation of a geometric
solution of V0,γ in the sense of Lemma 2.4 for every γ ∈ Γ. This condition is represented by the nonvanishing of a nonconstant polynomial
of degree at most 4Dγ3 for each γ ∈ Γ.
An algorithm for the computation of the minimal polynomial of u in
Vb described below can be extended to a computation of a geometric
solution of Vb in the sense of Lemma 2.4. This application of Lemma 2.4
141
requires that the coefficient vector of the linear form u does not annihilate a non-constant polynomial of degree at most 4D4 .
As a first step, we fix ρ ≥ 2. From Theorem 2.1 it follows that a linear
form u satisfying these conditions can be obtained by randomly choosing its
coefficients from the set {1, . . . , 6ρD4 } with error probability at most 1/ρ.
Second, we compute the monic minimal polynomial m
b u ∈ Q(T )[Y ] of
the linear form u in the curve Vb introduced in (5.12). For this purpose, we
approximate the Puiseux series expansions of the branches of Vb lying above
0 by means of a global Newton-Hensel lifting of the common zeros of the
zero-dimensional varieties V0,γ ⊂ An defined by the polynomials (5.14) for all
γ ∈ Γ (see Section 5.2.4).
This in turn requires the computation of a geometric solution of V0,γ for
(0)
every γ ∈ Γ. By means of a change of variables we take the system h1,γ =
(0)
0, . . . , hn,γ = 0 defining the variety V0,γ into a “diagonal” form (see Subsection
(0)
5.2.4 below), which allows us to compute the minimal polynomial mu,γ of
u in V0,γ . Since the linear form u satisfies condition (2) of the statement
of Lemma 2.4, from this procedure we derive an algorithm computing a
geometric solution of V0,γ according to Lemma 2.4.
Then we “lift” this geometric solution to a suitable (non-archimedean) approximation m
e γ of a factor mγ (over Q(T )) of the desired minimal polynomial
m
b u of u.
Q
In the third step we obtain the minimal polynomial m
b u = γ∈Γ mγ from
the approximate factors m
e γ , namely, we compute the dense representation of
the coefficients (in Q(T )) of m
b u , using Padé approximation (see Subsection
5.2.5 below). To conclude this step, we apply the proof of Lemma 2.4 to
derive an algorithm for computing a geometric solution of the variety Vb .
In the last step we compute a geometric solution of the variety V1 by
substituting 1 for T in the polynomials that form the geometric solution of
Vb .
This algorithm which solves the system h1 = 0, . . . , hn = 0 may be briefly
sketched as follows:
Algorithm 5.8.
Step I Choose the coefficients of a linear form u ∈ Q[X] at random from the
142
set {1, . . . , 6ρD4 }.
Step II For each γ ∈ Γ :
Step II.1 Find a geometric solution of the variety V0,γ defined in (5.15).
Step II.2 Obtain a straight-line program for the polynomials h1,γ , . . . , hn,γ
defined in (5.16) from the coefficients of h1 , . . . , hn and the entries
of γ ∈ Zn+1 .
Step II.3 “Lift” the computed geometric solution of V0,γ to an approximation
m
e γ of the factor mγ of m
b u by means of a symbolic Newton-Hensel
procedure.
Step III Obtain a geometric solution of the curve Vb :
Q
e γ of m
b u.
Step III.1 Compute the approximation m
e u := γ∈Γ m
Step III.2 Compute the dense representation of m
b u from m
e u using Padé approximation.
Step III.3 Find a geometric solution of Vb applying the proof of Lemma 2.4.
Step IV Substitute 1 for T in the polynomials which form the geometric solution
of Vb computed in the previous step to obtain a geometric solution of
the variety V1 .
Step I has been considered above. Next we discuss the following steps.
5.2.4.
Geometric solutions of the starting varieties
In this subsection we exhibit an algorithm which covers Step II.1 of Algorithm 5.8 and computes, for a given inner normal γ ∈ Γ, a geometric solution
(0)
of the variety V0,γ ⊂ (C∗ )n defined by the polynomials hi,γ (1 ≤ i ≤ n) for
polynomials h1 , . . . , hn satisfying assumptions (H1) and (H2). This algorithm
is based on the procedure presented in [HS95].
Fix a cell C = (C (1) , . . . , C (s) ) of type (k1 , . . . , ks ) of the given fine-mixed
subdivision of A and let γ ∈ Γ be its associated inner normal. For 1 ≤ ℓ ≤ s,
(0)
(0)
(ℓ)
(ℓ)
we denote by h1 , . . . , hkℓ the polynomials in the set {h1,γ , . . . , hn,γ } that are
supported in C (ℓ) . In the sequel, whenever there is no risk of confusion we
will not write the subscript γ indicating which cell we are considering.
(ℓ)
143
(ℓ)
Our hypotheses imply that h1 , . . . , hkℓ are Q-linear combinations of precisely kℓ + 1 monomials in Q[X] and, up to a multiplication by a monomial,
we may assume one of them to be the constant term. Denote these monof(ℓ) be the matrix of
mials by X αℓ,0 , . . . , X αℓ,kℓ , with αℓ,0 := 0 ∈ Zn . Let M
Qkℓ ×(kℓ +1) for which the following equality holds in Q[X, X −1 ]kℓ :
 α   (ℓ) 
h1
X ℓ,kℓ



.
(ℓ)
f  ..  =  ... 
(5.27)
M
,
αℓ,0
(ℓ)
X
hk
ℓ
and let M(ℓ) denote the square (kℓ × kℓ )-matrix obtained by deleting the last
f(ℓ) . Set
column from M


M(1)
0
···
0
 0
M(2) · · ·
0 


M := 
,
.
..
 0
0
0 
0
0
· · · M(s)
where 0 here represents different block matrices with all its entries equal to
0 ∈ Q. Then M is the matrix defined by the coefficients of the nonconstant
(0)
(0)
terms of the (Laurent) polynomials h1,γ , . . . , hn,γ , up to a translation.
Due to condition (H2) we have that the matrix M is invertible, which
in turn implies that the square matrices M(ℓ) are invertible for 1 ≤ ℓ ≤ s.
f(ℓ) for
Following [HS95], we apply Gaussian elimination to the matrix M
1 ≤ ℓ ≤ s and obtain a set of kℓ + 1 binomials
 α
 


1 0 0 . . . −cαℓ,kℓ
X αℓ,kℓ − cαℓ,kℓ
X ℓ,kℓ
0 1 0 . . . −cα
 X αℓ,kℓ −1  X αℓ,kℓ −1 − cα

ℓ,kℓ −1 
ℓ,kℓ −1  
 

=

 
.

..
..
..

 ..
 

.
.
.
αℓ,0
αℓ,0
X
0 0 ... 1
−cαℓ,0
X
− cαℓ,0
that generate the same linear subspace of Q[X, X −1 ] as the polynomials in
(5.27). Therefore, for 1 ≤ ℓ ≤ s the set of common zeros in (C∗ )n of the
(ℓ)
(ℓ)
polynomials h1 , . . . , hkℓ is given by the system X αℓ,kℓ = cαℓ,kℓ , . . . , X αℓ,1 =
cαℓ,1 . Putting these s systems together, we obtain a binomial system defining
V0,γ of the form
(5.28)
X α1 = p 1 , . . . , X αn = p n ,
144
with αi ∈ Zn and pi ∈ Q \ {0} (1 ≤ i ≤ n). Note that the second part of
condition (H2) ensures the non-vanishing of the constants pi for 1 ≤ i ≤ n.
Now, let E denote the (n × n)-matrix whose columns are the exponent
vectors α1 , . . . , αn . Using [Sto00, Proposition 8.10], we obtain unimodular
matrices K = (ki,j )1≤i,j≤n , L = (li,j )1≤i,j≤n of Zn×n , and a diagonal matrix diag(r1 , . . . , rn ) ∈ Zn×n which give the Smith Normal Form for E, i.e.,
matrices such that the identity
K · E · L = diag(r1 , . . . , rn )
(5.29)
holds in Zn×n . We observe that the upper bound
log kKk ≤ (4n + 5)(log n + log kEk)
(5.30)
holds, where kAk denotes the maximum of the absolute value of the entries
of a given matrix A [Sto00, Proposition 8.10].
Let Z1 , . . . , Zn be new indeterminates, and write Z := (Z1 , . . . , Zn ). We
k
k
introduce the change of coordinates given by Xi := Z1 1,i · · · Znn,i for 1 ≤ i ≤
n. Making this change of coordinates in (5.28) we obtain the system
Z Kα1 = p1 , . . . , Z Kαn = pn ,
which is equivalent to the “diagonal” system
r
Zj j
=
n
Y
i=1
(Z
Kαi li,j
)
=
n
Y
l
pii,j =: qj
(1 ≤ j ≤ n).
i=1
Inverting some of the coefficients qj if necessary we may assume without loss
of generality that the integers r1 , . . . , rn are positive. We have thus a very
convenient description of the variety V0,γ by a diagonal polynomial system
in the coordinate system of An defined by Z1 , . . . , Zn . We shall compute a
geometric solution of V0γ in such a coordinate system, which will be then
used to compute a geometric solution of V0,γ in the “standard” coordinate
system defined by X1 , . . . , Xn .
Example. We illustrate the above procedure for the variety V0,γ (3) of (5.22),
namely,
( (0)
h1,γ (3) = 1 − X12 X22 ,
V0,γ (3) = {(1, −1), (−1, 1), (i, −i), (−i, i)}.
(0)
h2,γ (3) = X12 X2 + X1 X22 ,
(5.31)
145
Here the binomial system in (5.28) and the corresponding exponent vector
matrix E are
(
X12 X22 = 1,
2 1
.
and
E=
2 −1
X1 X2−1 = −1,
1 0
0 1
1 0
, and
, we get K · E · L =
and L :=
Taking K :=
0 4
1 −2
1 1
hence, making the change of coordinates X1 = Z1 Z2 , X2 = Z2 we obtain the
equivalent diagonal system
(
Z1 = −1,
(5.32)
Z24 = 1.
The algorithm for computing a geometric solution of the variety V0,γ in
the coordinate system defined by Z1 , . . . , Zn takes as input the set of polynomials Z1r1 − q1 , . . . , Znrn − qn ∈ Q[Z1 , . . . , Zn ] and outputs a linear form
u
e ∈ Q[Z1 , . . . , Zn ] which separates the points of V0,γ , the minimal polynomial mue ∈ Q[Y ] of u
e in V0,γ and the parametrizations w
e1 , . . . , w
en of Z1 , . . . , Zn
by the zeros of mue . Observe that Dγ = r1 · · · rn = #(V0,γ )
For this purpose we apply the algorithm for solving a “diagonal” system
underlying Lemma 3.11. Combining Lemma 3.11 and Proposition 3.14 we
obtain the following result.
Proposition 5.9. Suppose that the coefficients of the linear form u
e are randomly chosen in the set {1, . . . , 4nρDγ3 }, where ρ is a fixed positive integer.
Then the algorithm described above computes a geometric solution of the variety V0,γ (in the coordinate
system Z1 , . . . , Zn ) with error probability at most
2
1/ρ using O nM(Dγ ) arithmetic operations in Q.
Example. For the system (5.32) defining the variety V0,γ (3) in the coordinate
system Z1 , Z2 , taking the separating linear form u
e = Z1 + Z2 we obtain:
m1 := Y + 1,
m2 := ResYe ((Y − Ye )4 − 1, Ye + 1) = Y 4 + 4Y 3 + 6Y 2 + 4Y .
146
Then, the minimal polynomial of u
e is mue = Y 4 + 4Y 3 + 6Y 2 + 4Y .
Next we compute a geometric solution of the variety V0,γ in the original
coordinate system defined by X1 , . . . , Xn . As before, we consider γ ∈ Γ fixed.
For this purpose, we compute the minimal polynomial mu ∈ Q[Y ] of a
linear form u = u1 X1 + · · · + un Xn ∈ Q[X1 , . . . , Xn ] in V0,γ . Let V0,γ :=
QDγ
(D ,γ)
(j,γ)
(1,γ)
(Y − u(x0 )). In order to
{x0 , . . . , x0 γ }. Then we have mu (Y ) = j=1
compute mu , we use the polynomials mue , w
e1 , . . . , w
en which form the previously computed geometric solution of V0,γ in the variables Z1 , . . . , Zn : from
k
k
the identities Xi := Z1 1,i · · · Znn,i (1 ≤ i ≤ n) we deduce that mu equals the
minimal polynomial of the image of the projection ηu : V0,γ → A1 defined
P
(γ)
ei (e
u),
by ηu (z1 , . . . , zn ) := ni=1 ui z1k1,i · · · znkn,i . Now, the identities Zi = w
which hold in Q[V0,γ ] for 1 ≤ i ≤ n, imply that
u=
n
X
i=1
e1 (e
u)
ui w
k1,i
··· w
en (e
u)
kn,i
(5.33)
holds in Q[V0,γ ], from which we easily conclude that mu satisfies the following
identity:
!
n
X
kn,i
k1,i
, mue (Ye ) . (5.34)
··· w
en (Ye )
e1 (Ye )
ui w
mu (Y ) = Res e Y −
Y
i=1
Example. We compute now a geometric solution of V0,γ (3) in the coordinate
system X1 , X2 for the linear form u = X1 − X2 from its geometric solution
in the coordinates Z1 , Z2 (see Subsection 5.2.4):
mue = Y 4 + 4Y 3 + 6Y 2 + 4Y ,
w
e1 = −1,
w
e2 = Y + 1.
From the change of coordinates X1 = Z1 Z2 , X2 = Z2 leading to system
(5.32), we have u = Z1 Z2 − Z2 and hence u = −2(e
u + 1). Therefore,
mu = ResYe (Y +2(Ye +1)), Ye 4 +4Ye 3 +6Ye 2 +4Ye ) = Y 4 −16.
147
Now,
complexity of this step. We compute the monomials
k we estimate the
kn,i
1,i
w
e1 (e
u)
··· w
en (e
u)
(1 ≤ i ≤ n) in the
right-hand side of (5.33) modulo
2
mue (Y ), with O n log(maxi,j |ki,j |)M(Dγ ) arithmetic operations in Q. From
(5.30) it follows that
O n2 log(max |ki,j |)M(Dγ ) = O n3 log(nkEγ k)M(Dγ ) ,
i,j
where Eγ is the matrix of the exponents of the cell corresponding to the inner
normal γ. Observe that all these steps are independent of the coefficients of
the linear form u we are considering and therefore do not introduce any
division by a nonconstant polynomial in the coefficients u1 , . . . , un .
In the next
step we compute the right-hand side of (5.33) modulo mue (Y ),
with O nDγ arithmetic operations in Q. Then we compute the resultant
(5.34) by a process which interpolates (5.34) in the variable Y to reduce the
question to the computation of Dγ + 1 univariate resultants. This requires
O M(Dγ )2 arithmetic operations in Q.
If the linear form u separates the points of V0,γ , then we can extend the
algorithm for computing mu (Y ) to an algorithm for computing a geometric
solution of V0,γ with the algorithm underlying the proof of Lemma 2.4. This
extension requires that the coefficients u1 , . . . , un of the linear form u do not
annihilate the denominators in Q[Λ] which arise from the application of the
algorithm described above to the generic version Λ1 X1 + · · · + Λn Xn of the
linear form u. Such denominators arise only during the computation of the
generic version of the resultant (5.34). Hence, with a similar analysis as in
the proof of Proposition 5.9, we conclude that, if the coefficients of u are
chosen randomly in the set {1, . . . , 4nρDγ3 }, then the error probability of our
algorithm is bounded by 1/ρ. In conclusion, we have the following result.
Proposition 5.10. Suppose that we are given a geometric solution of V0,γ
in the coordinate system Z1 , . . . , Zn , as provided by the algorithm underlying
Proposition 5.9, and the coefficients of the linear form u are randomly chosen in the set {1, . . . , 4nρDγ3 }, where ρ is a fixed positive integer. Then the
algorithm described above computes a geometric solution of the variety V0,γ
with error probability at most 1/ρ using O n3 log(nkEγ k)M(Dγ )2 arithmetic
operations in Q.
Finally, from Propositions 5.9 and 5.10 and the fact that kEγ k ≤ 2Q holds
for Q := max1≤i≤n {kqk; q ∈ ∆i }, we immediately deduce the following result.
148
Theorem 5.11. Suppose that the coefficients of the linear forms u
e and
u of the statement of Propositions 5.9 and 5.10 are chosen at random in
the set {1, . . . , 4nρD3 }, where ρ is a fixed positive integer. Then the algorithm underlying Propositions 5.9 and 5.10 computes a geometric solution
of the varieties V0,γ for all γ ∈ Γ with error probability at most 2/ρ using
O n3 log(nQ)M(D)2 arithmetic operations in Q.
Example. For the polynomial system (5.17) we are considering and the linear
form u = X1 − X2 , the first step of Algorithm 5.8 computes the following
geometric solutions of the varieties of (5.20), (5.21) and (5.22) respectively,
as explained above:
V0,γ (1) = {(−1, −y − 1) ∈ C2 : y 2 + 2y = 0},
V0,γ (2) = {((y − 1, −1) ∈ C2 : y 2 − 2y = 0},
V0,γ (3) = {( 21 y, − 12 y) ∈ C2 : y 4 − 16 = 0}.
5.2.5.
(5.35)
A geometric solution of the curve Vb
We recall the definition of the variety Vb . Let I denote the ideal of Q[X, T ]
generated by the polynomials
X
b
hi (X, T ) =
ci,q X q T ωi (q) (1 ≤ i ≤ n)
q∈∆i
of (5.11), which form the polyhedral deformation of the generic polynomials
h1 , . . . , hn , and let J denote the Jacobian determinant of b
h1 , . . . , b
hn with
respect to the variables X1 , . . . , Xn . Let V (I) be the set of common zeros in
An+1 of b
h1 , . . . , b
hn . Then Vb := V (I : J ∞ ).
Alternatively, let π : V (I) → A1 be the linear projection defined by
π(x, t) := t. Consider
comSr+s the decomposition of V (I) into its irreducible
1
ponents V (I) = i=1 Ci . Suppose that the restriction π|Ci : Ci → A of the
projection π is dominant for
S1 ≤ i ≤ r and is not dominant for r + 1 ≤ i ≤ s.
We shall show that Vb := ri=1 Ci holds, i.e., Vb is the union of all the irreducible components of V (I) which project dominantly over A1 . Furthermore,
we shall show that Vb ⊂ An+1 is a curve which constitutes a suitable deformation of the variety defined by the system h1 = 0, . . . , hn = 0. For this
purpose, adapting Proposition 3.3 we deduce the following technical lemma.
149
Lemma 5.12. Let F1 , . . . , Fn ∈ Q[X, T ] be polynomials which generate an
ideal I := (F1 , . . . , Fn ) ⊂ Q[X, T ] and let J denote the Jacobian determinant
of F1 , . . . , Fn with respect to the variables X. Set V := {(x, t) ∈ An+1 :
F1 (x, t) = 0, . . . , Fn (x, t) = 0} and consider the linear projection π : V → A1
defined by π(x, t) := t. Assume that #π −1 (t) ≤ D holds for generic values
of t ∈ A1 and that there exists a point t0 ∈ A1 such that the fiber π −1 (t0 ) is
a zero-dimensional variety of degree D with J(x, t0 ) 6= 0 for every (x, t0 ) ∈
π −1 (t0 ).
Let Vdom be the union of all the irreducible components C of V with π(C) =
A1 . Then:
Vdom is a nonempty equidimensional variety of dimension 1.
Vdom is the union of all the irreducible components of V having a nonempty intersection with π −1 (t0 ).
Vdom = V (I : J ∞ ).
The restriction π|Vdom : Vdom → A1 is a dominant map of degree D.
Now we return to the study of the variety Vb and show that the assumptions of Lemma 5.12 hold. Observe that π −1 (t) = Vt × {t} holds for every
t ∈ A1 , where Vt := {x ∈ An : b
h1 (x, t) = 0, . . . , b
hn (x, t) = 0}. Furthermore,
the polynomials b
h1 (X, t), . . . , b
hn (X, t) are obtained by a suitable substitution of the variables Ω of the generic polynomials H1 , . . . , Hn ∈ Q[Ω, X]
with supports ∆1 , . . . , ∆n introduced in (5.26). Indeed, if c = (c1 , . . . , cn )
is the vector of coefficients of h1 , . . . , hn , the coefficient vector of b
hi (X, t)
(1 ≤ i ≤ n) is (ci,q tωi (q) )q∈∆i for every t ∈ A1 . By Lemma 5.4, there exists a nonzero polynomial P (0) ∈ Q[Ω] such that, for any c′ = (c′1 , . . . , c′n )
with P (0) (c′ ) 6= 0, the associated sparse system defines a zero-dimensional
variety. In particular, the coefficients c = (c1 , . . . , cn ) of our input polynomials h1 := H1 (c1 , X), . . . , hn = Hn (cn , X) satisfy P (0) (c) 6= 0. This shows
(0)
that the polynomial PT ∈ Q[T ] obtained by substituting Ωi,q 7→ ci,q T ωi (q)
(1 ≤ i ≤ n, q ∈ ∆i ) in the polynomial P (0) is nonzero, since it does not
vanish at T = 1. We conclude that Vt is a zero-dimensional variety for all
but a finite number of t ∈ A1 . Thus, π −1 (t) is finite for generic values of
t ∈ A1 .
150
Finally, by condition (H1), the fiber π −1 (1) = V (h1 , . . . , hn ) × {1} is a
zero-dimensional variety of degree D = deg(π) and the Jacobian determinant
J := det(∂ b
hi /∂Xj )1≤i,j≤n does not vanish at any of its points. On the other
hand, the fact that #π −1 (t) ≤ D holds for generic values t ∈ A1 follows from
the BKK theorem.
This shows that the variety V (I) and its defining polynomials b
h1 , . . . , b
hn
satisfy all the assumptions of Lemma 5.12. Thus, we have the following
result.
Lemma 5.13. The variety Vb ⊂ An+1 is a curve. Furthermore, every irreducible component of Vb has a nonempty intersection with the fiber π −1 (1) of
the projection map π : Vb → A1 .
Generic linear projections of Vb .
In order to compute a geometric solution of the space curve Vb , we shall
first exhibit a procedure for computing the minimal polynomial of a generic
linear projection of Vb . Let u ∈ Q[X1 , . . . , Xn ] be a linear form which separates the points of the “initial varieties” V0,γ for all the inner normals γ :=
(γ1 , . . . , γn+1 ) of the lower facets of the polyhedral deformation under consideration. Let πu : Vb → A2 be the morphism defined by πu (x, t) := (t, u(x)).
Since the projection map π : Vb → A1 defined by π(x, t) := t is dominant, it
follows that the Zariski closure of the image of πu is a Q-definable hypersurface of A2 . Denote by Mu ∈ Q[T, Y ] a minimal defining polynomial for this
hypersurface. For the sake of the argument, we shall assume further that the
identity deg(π) = D, and thus degY Mu = D, holds.
We can apply estimate (5.4) of Lemma 5.3 in order to estimate degT Mu
in combinatorial terms (compare with [PS08, Theorem 1.1]). Indeed, let
b1 , . . . , Q
bn ⊂ Rn+1 be the Newton polytopes of the polynomials b
Q
h1 , . . . , b
hn
n+1
of (5.11), and let ∆ ⊂ R
be the standard n–dimensional simplex in the
hyperplane {T = 0}. Then the following estimate holds:
b1 , . . . , Q
bn ).
degT Mu ≤ E := M(∆, Q
(5.36)
Furthermore, equality holds in (5.36) for a generic choice of the coefficients
of the polynomials b
hi and the linear form u.
151
Our purpose is to exhibit a procedure for computing the unique monic
multiple m
b u in Q(T )[Y ] of Mu of degree D. This polynomial can be alternatively defined in terms of the Puiseux series solutions to the polynomials
b
h1 , . . . , b
hn as we explain in what follows.
Since the projection map π : Vb → A1 is dominant, it induces an extension
Q[T ] ֒→ Q[Vb ], where Q[Vb ] denotes the coordinate ring of Vb . This variety
being a curve, Q[Vb ] turns out to be a finitely generated Q[T ]-module. Thus,
tensoring with Q(T ), we deduce that Q[Vb ] ⊗ Q(T ) is a Q(T )-vector space
of finite dimension. We claim that Q[Vb ] ⊗ Q(T ) = Q[V (I)] ⊗ Q(T ) holds.
Indeed, since Vb consists of the irreducible components of V (I) which are
mapped dominantly onto A1 by the projection π, for each of the remaining
irreducible components C of V (I), the set π(C) ⊂ C is a zero-dimensional
Q-definable variety. This implies that I(C) ∩ Q[T ] 6= {0} holds.
Let m
b u be the minimal polynomial of u in the extension Q(T ) ֒→ Q[Vb ] ⊗
Q(T ). The fact that Q[Vb ] ⊗ Q(T ) is finite-dimensional Q(T )-vector space
shows that the affine variety V := {x̄ ∈ ASn (Q(T )∗ ) : b
h1 (x̄) = 0, . . . , b
hn (x̄) =
∗
1/q
0} has dimension zero. Here Q(T ) := q∈N Q((T )) denotes the field of
Puiseux series in the variable T over Q and b
h1 , . . . , b
hn are considered as
elements of Q(T )[X]. Our hypotheses imply that there exist D distinct
(ℓ)
(ℓ)
n-tuples x(ℓ) := (x1 , . . . , xn ) ∈ (Q(T )∗ )n of Puiseux series such that the
following equalities hold in Q(T )∗ for 1 ≤ ℓ ≤ D:
b
h1 (x(ℓ) , T ) = 0 , . . . , b
hn (x(ℓ) , T ) = 0
(5.37)
(see [HS95]). Since Q[Vb ]⊗Q(T ) is the coordinate ring of the Q(T )-variety V,
from (5.37) we deduce that the dimension of Q[Vb ] ⊗ Q(T ) over Q(T ) equals
D. Moreover, since degY m
b u = D holds as a consequence of our assumptions,
we conclude that
D
Y
m
bu =
Y − u(x(ℓ) ) .
(5.38)
ℓ=1
Since Mu (T, u(X)) ∈ I(Vb ), it follows that Mu (T, u(X)) = 0 holds in Q[Vb ] ⊗
Q(T ), from which we conclude that Mu is a multiple of m
b u by a factor in
Q(T )[Y ]. Taking into account that both are polynomials of degree D in the
variable Y and that m
b u is monic in this variable, we deduce that m
b u is the
quotient of Mu by its leading coefficient. We summarize our arguments in
the following statement.
152
Lemma 5.14. Let πu : Vb → A2 be the projection defined by πu (x, t) :=
t, u(x) . Assume that the identity deg(π) = D holds and let Mu ∈ Q[T, Y ]
2 ). Denote by m
bu
be the minimal defining polynomial of the hypersurface πu (AQ
D
(ℓ)
the only monic multiple of Mu in Q(T )[Y ]. Then m
b u (Y ) = ℓ=1 (Y −u(x )),
where x(1) , . . . , x(D) ∈ An (Q(T )∗ ) are the solutions of (5.37).
Next, we group the roots u(x(ℓ) ) of the polynomial m
b u according to the
facet from where they arise. With notations as in Section 5.2.1, let Γ ⊂ Zn+1
be the set of primitive integer vectors of the form γ := (γ1 , . . . , γn , γn+1 ) ∈
Zn+1 with γn+1 > 0 for which there is a cell C = (C (1) , . . . , C (s) ) of type
b has inner normal
(k1 , . . . , ks ) of the subdivision of A induced by ω such that C
b
γ. As asserted in Section 5.2.1, if γ ∈ Γ is the inner normal of the lifting C
of a cell C of type (k1 , . . . , ks ), there exist Dγ := k1 ! · · · ks ! · Vol(C) vectors
(j,γ)
(j,γ)
of Puiseux series x(j,γ) := (x1 , . . . , xn ) ∈ An (Q(T )∗ ) (1 ≤ j ≤ Dγ ) of the
form
X (j,γ) γi +m
(j,γ)
xi
:=
xi,m T γn+1
m≥0
satisfying (5.37). Considering the projection of the branches of Vb parametrized
by the Dγ vectors of Puiseux series x(j,γ) for each γ ∈ Γ, we obtain the following element mγ of Q((T 1/γn+1 ))[Y ]:
mγ :=
Dγ
Y
j=1
Y − u(x(j,γ) ) .
(5.39)
From (5.2) we conclude that (5.38) may be expressed in the following way:
m
bu =
Y
mγ .
(5.40)
γ∈Γ
Since m
b u belongs to Q(T )[Y ] and its primitive multiple Mu ∈ Q[T, Y ]
satisfies the degree estimate degT Mu ≤ E, in order to compute the dense
representation of m
b u we shall compute the Puiseux expansions of the coefficients of the factors mγ ∈ Q((T 1/γn+1 ))[Y ] of m
b u truncated up to order 2E.
Using Padé approximation it is possible to recover the dense representation
of m
b u from this data.
153
(j,γ)
(j,γ)
(j,γ)
Fix γ ∈ Γ and set xm := (x1,m , . . . , xn,m ) for every m ≥ 0 and 1 ≤ j ≤
Dγ . Since
X
γ1 +m
X
γn +m
(j,γ)
(j,γ) γn+1
b
x1,m T γn+1 , . . . ,
hi
xn,m
T
,T = 0
m≥0
m≥0
holds for 1 ≤ j ≤ Dγ and 1 ≤ i ≤ n, we have
P
(j,γ) γn +m
(j,γ) γ1 +m
γn+1
T
,
T
x
x
T
,
.
.
.
,
n,m
m≥0
m≥0 1,m
P
P
(j,γ) m
(j,γ) m
γn+1
γn
γ1
−mi b
= T
hi T
m≥0 xn,m T , T
m≥0 x1,m T , . . . , T
P
(j,γ) m
= hi,γ
m≥0 xm T , T ,
P
0 = T −mi b
hi
according to (5.16). Therefore the polynomial mγ (T γn+1 , Y ) ∈ Q((T ))[Y ] can
be expressed in terms of the power series solutions
(j,γ)
σ (j,γ) := (σ1
, . . . , σn(j,γ) ) :=
X
m
x(j,γ)
m T
(1 ≤ j ≤ Dγ )
(5.41)
m≥0
of h1,γ , . . . , hn,γ . Indeed, from (5.39) it follows that
mγ (T γn+1 , Y ) =
=
=
=
where uγ :=
Pn
i=1
QDγ
j=1
QDγ
j=1
QD γ
j=1
QDγ
j=1
P
(j,γ)
ui m≥0 xi,m T γi +m
P
P
(j,γ)
Y − m≥0 ni=1 ui xi,m T γi T m
P
(j,γ)
Y − m≥0 uγ (xm )T m
P
(j,γ)
Y − uγ ( m≥0 xm T m ) =: muγ (T, Y ),
Y −
Pn
i=1
ui T γi Xi . In conclusion, we have the following result.
Lemma 5.15. Fix γ := (γ, . . . , γn+1 ) ∈ Γ and let mγ be as in (5.39). Then
the Laurent polynomial mγ (T γn+1 , Y ) ∈ Q((T ))[Y ] equals
P the minimal polynomial muγ (T,Y ) of the projection induced by uγ := ni=1 ui T γi Xi on the subvariety of An (Q(T )∗ ) consisting of the set of power series {σ (1,γ) , . . . , σ (Dγ ,γ) }
of (5.41).
This lemma will be critical in order to obtain suitable approximations to
the Laurent polynomials mγ (T γn+1 , Y ) in Q((T ))[Y ].
154
A procedure for computing m
b u.
Now we exhibit a procedure for computing the minimal polynomial m
b u,
which is based on the computation
of
the
Laurent
polynomials
m
arising
in
γ
Q
the factorization of m
b u = γ∈Γ mγ in terms of Puiseux expansions according
to Lemmas 5.14 and 5.15. Then we will apply Lemma 2.4 to this procedure
in order to obtain an algorithm for computing a geometric solution of the
curve Vb .
In order to describe this approximation, we recall the following termie ∈ Q((T )) and s ∈ Z, we say that G
e approximates G with
nology: for G, G
e has order at least s + 1 in T .
precision s in Q((T )) if the Laurent series G − G
e mod (T s+1 ). Furthermore, if G, G
e are two
We shall use the notation G ≡ G
e approximates G with
elements of a polynomial ring Q((T ))[Y ], we say that G
e approximates the correspondprecision s if every coefficient e
a ∈ Q((T )) of G
ing coefficient a ∈ Q((T )) of G with precision s (in the sense of the previous
definition).
Fix γ := (γ1 , . . . , γn ) ∈ Γ. Let Vγ := {σ (j,γ) : 1 ≤ j ≤ Dγ } denote the
subset of V underlying γ. In order to compute the required approximation
of the polynomial muγ of the statement of Lemma 5.15, we first compute
a corresponding approximation of the polynomials that form a geometric
solution of the variety Vγ . Observe that
(j,γ)
{σ (j,γ) (0) : 1 ≤ j ≤ Dγ } = {x0
: 1 ≤ j ≤ Dγ }
(0)
∗ n
= V (h1,γ , . . . , h(0)
n,γ ) ∩ (C )
= V (h1,γ (X, 0), . . . , hn,γ (X, 0)) ∩ (C∗ )n = V0,γ
(j,γ)
holds. Since det(∂hi,γ (X, 0)/∂Xk )1≤i,k≤n (x0 ) 6= 0 holds for 1 ≤ j ≤ Dγ ,
we may apply of the global Newton iterator of [GLS01] in order to “lift” the
given geometric solution of V0,γ to the geometric solution of the variety Vγ
associated to the linear form u ∈ Q[X] with any prescribed precision.
(0)
(0)
(0)
Suppose that we are given polynomials mu,γ , wu,1,γ , . . . , wu,n,γ ∈ Q[Y ]
which form a geometric solution of V0,γ , as provided by the algorithm underly(j) (j) (j,γ)
(0)
(0)
ing Theorem 5.11. Recall that mu,γ u(x0 ) = 0 and (x0 )i = wu,i,γ u(x0 )
hold for 1 ≤ i ≤ n and 1 ≤ j ≤ Dγ . The global Newton iterator is a recursive
(k)
(k)
(k)
procedure whose kth step computes approximations mu,γ , wu,1,γ , . . . , wu,n,γ ∈
155
Q[T, Y ] of the polynomials mu,γ , wu,1,γ , . . . , wu,n,γ which form the geometric
solution of Vγ associated with the linear form u with precision 2k for any
k ≥ 0.
We may assume without loss of generality that γi ≥ 0 and 0 = min{γ1 , . . . ,
γn } hold for 1 ≤ i ≤ n. Indeed, if there exists γi < 0, setting γi0 :=
min{γ1 , . . . , γn } we have
T
−γi0 Dγ
mγ (T
γn+1
,T
γi 0
Y) =
=
=
Dγ
Y
T
j=1
Dγ Y
j=1
Dγ Y
j=1
−γi0
T
Y −T
γi0
Y −
n
X
i=1
−γi0
n
X
ui
i=1
Y −
ui
n
X
i=1
ui
X
m≥0
X
X
(j,γ)
xi,m T γi +m
m≥0
(j,γ)
xi,m T γi +m
m≥0
(j,γ)
xi,m T γi −γi0 +m .
(5.42)
Since γi − γi0 ≥ 0 holds for 1 ≤ i ≤ n, this shows that the computation of an
approximation of muγ = mγ (T γn+1 , Y ) can be easily reduced to a situation
in which γi ≥ 0 holds for 1 ≤ i ≤ n.
Note that the global Newton iterator cannot be directly applied in order
to compute the geometric solution of {σ (j,γ) ; 1 ≤ j ≤ Dγ } associated with
the linear form uγ ∈ Q[T ][X], because the coefficients of uγ are nonconstant
polynomials of Q[T ]. Indeed, two critical problems arise:
1. Although by hypothesis uγ separates the points of Vγ , it might not
separate the points of V0,γ and it is not clear from which precision on,
the corresponding approximations of the points of Vγ are separated by
uγ . Requiring uγ to be a separating form for all the approximations of
the points of Vγ is an essential hypothesis for the iterator of [GLS01]
which cannot be suppressed without causing a significant growth of the
complexity of the procedure (see [Lec02], [Lec03]).
2. The iterator of [GLS01] makes critical use of the fact that the coefficients of the linear form under consideration are elements of Q in order
to determine how a given precision can be achieved.
Nevertheless, we shall exhibit a modification of the procedure which computes
156
an approximation of muγ (T, Y ) with precision 2γn+1 E without changing the
asymptotic number of arithmetic operations performed.
In order to circumvent (1) we require anPadditional generic condition
on the coefficients u1 , . . . , un defining uγ := ni=1 ui T γi Xi , namely, the uγ
separates the first Mγ terms of the series σ (j,γ) . Our next result asserts
that for a random choice of the coefficients uγ , this condition is likely to be
satisfied.
Lemma 5.16. For a random choice of values u1 , . . . , un in the set {1, . . . , ρDγ2 },
P
P γ (j,γ) m
the linear form uγ := ni=1 ui T γi Xi separates the initial terms M
m=0 xm T
(j,γ)
of the power series σ
(1 ≤ j ≤ Dγ ) with probability at least 1 − 1/ρ, where
Mγ := max{γ1 , . . . , γn }.
P
Proof. For a given linear form uγ := ni=1 ui T γi Xi as in the statement of the
Pn
P
(j,γ) m
u
x
T for every 1 ≤ j ≤
lemma, we have uγ (σ (j,γ) ) = m≥0
i
i,m−γ
i=1
i
(j,γ)
Dγ , where xi,m−γi := 0 for m < γi . We make the following claim.
Claim 5.17. Let Λ1 , . . . , Λn be indeterminates over C[T,X]. Then the following inequality holds for every 1 ≤ j, h ≤ Dγ with j 6= h:
Mγ n
X
X
m=0
(j,γ)
Λi xi,m−γi
i=1
T
m
6=
Mγ n
X
X
m=0
(h,γ)
Λi xi,m−γi
i=1
T m.
Proof ofPClaim. Suppose
on the
there exist j 6= h such that
PMcontrary
Pn that(h,γ)
P
(j,γ) m
Mγ
n
γ
m
−γi
Λi
= m=0
i=1 Λi xi,m−γi T . Substituting T
m=0
i=1 Λi xi,m−γi T
P Mγ P n
(j,γ)
m−γi
for Λi in this identity for i = 1, . . . , n, we have m=0 i=1 Λi xi,m−γi T
=
P Mγ P n
(h,γ)
m−γi
, that is
m=0
i=1 Λi xi,m−γi T
γ −γi
n MX
X
i=1
(j,γ)
Λi xi,m T m
=
m=0
γ −γi
n MX
X
i=1
(h,γ)
Λi xi,m T m .
m=0
Substituting 0 for T in this identity, we deduce that
n
X
i=1
(j,γ)
Λi xi,0
=
n
X
(h,γ)
Λi xi,0 ,
i=1
(j,γ)
(j,γ)
(j,γ)
which contradicts the fact that the vectors x0
= (x1,0 , . . . , xn,0 ) (1 ≤
j ≤ Dγ ) are all distinct. This finishes the proof of the claim.
157
m
P γ Pn
(j,γ)
(h,γ)
By the claim we see that the polynomial M
i=1 Λi (xi,m−γi −xi,m−γi ) T
m=0
of Q[Λ][T ] is nonzero, and therefore has a nonzero coefficient
Q aj,h ∈ C[Λ] for
every 1 ≤ j < h ≤ Dγ . Consider the polynomial Aγ (Λ) := 1≤j<h≤Dγ aj,h ∈
C[Λ]. Since aj,h has degree
1 for every 1 ≤ j < h ≤ Dγ , it follows
. . , un ) ∈ Cn with
that A has degree D2γ . Furthermore, for every (u1 , .P
Aγ (u1 , . . . , un ) 6= 0, the corresponding polynomial uγ := ni=1 ui T γi Xi sepaPMγ (j,γ) m
xm T of the power series σ (j,γ) (1 ≤ j ≤ Dγ ).
rates the initial terms m=0
Therefore, by Theorem 2.1 we see that, for a random choice of the coefficients
u1 , . . . , un in the set {1, . . . , ρDγ2 }, the linear form uγ separates the first Mγ
terms of the points of Vγ with probability at least 1 − 1/ρ.
Assume that the coefficients u1 , . . . , un satisfy the statement of the lemma.
A straight-line program for computing the polynomials h1,γ , . . . , hn,γ can be
easily derived from one computing h1 , . . . , hn and the coordinates of γ. We
assume this given, Step II.2 completed, and defer its complexity analysis. The
algorithm for computing an approximation of muγ (Step II.3) consists of the
following three steps:
(Step II.3.a) We compute a suitable approximation
Pnto the geometric solution of Vγ associated to the linear form u :=
i=1 ui Xi by means of
κ0 := ⌈log(Mγ + 1)⌉ steps of the global Newton iterator of [GLS01].
(Step II.3.b) We use the approximation of the previous step in order to obtain
(κ )
(κ )
(κ )
a corresponding approximation muγ0 , wuγ0,1 , . . . , wuγ0,n of the polynomials that
form the geometric solution of Vγ associated with uγ .
(Step II.3.c) We apply an adaptation of the global Newton iterator which
(κ )
(κ )
(κ )
takes as input the polynomials of the previous step muγ0 , wuγ0,1 , . . . , wuγ0,n and
outputs the required approximation to the polynomials muγ , wuγ ,1 , . . . , wuγ ,n
that form the geometric solution of Vγ associated with uγ .
Proposition 5.18. Fix γ := (γ1 , . . . , γn ) ∈ Γ and assume that a geometric solution of the variety V0,γ is given, as provided by Theorem 5.11. Assume further that the coefficients of the linear form u of the given geometric
solution of V0,γ are randomly chosen in the set {1, . . . , 4ρDγ3 } for a given
ρ ∈ N. Then the algorithm above computes an approximation to the polynomial muγ ∈ Q((T ))[Y ] with precision 2Eγn+1 . The procedure requires
158
O (nLγ + n3 )M(Dγ ) M(Mγ )M(Dγ )/ log(Mγ ) + M(Eγn+1 ) arithmetic operations in Q , where Mγ := max{γ1 , . . . , γn } and Lγ is the number of arithmetic operations required to evaluate the polynomials hi,γ of (5.16), and has
error probability at most 2/ρ.
Proof. We consider Steps II.3.a, II.3.b, II.3.c in detail. Step II.3.a takes as in(0)
(0)
(0)
put the given geometric solution mu,γ , wu,1,γ , . . . , wu,n,γ of V0,γ , and performs
κ0 := ⌈log(Mγ + 1)⌉ times the global Newton iterator of Theorem 2.2 to
(κ0 )
(κ )
(κ0 )
∈ Q[T, Y ] such that the following
obtain polynomials mu,γ0 , wu,1,γ
, . . . , wu,n,γ
conditions hold:
(i)u,κ0
(ii)u,κ0
(iii)u,κ0
(iv)u,κ0
(j,γ)
(κ )
(κ )
(κ )
(κ )
degY mu,γ0 = Dγ and degT mu,γ0 ≤ Mγ ,
0
0
degY wu,i,γ
< Dγ and degT wu,i,γ
≤ Mγ for 1 ≤ i ≤ n,
QDγ
(j,γ)
(κ )
mod (T Mγ +1 ),
Y − ϕ κ0
mu,γ0 ≡ j=1
(j,γ)
(κ0 )
(j,γ) σi
≡ wu,i,γ
mod (T Mγ +1 ) for 1 ≤ i ≤ n.
T, ϕκ0
Here ϕκ0 is the Taylor expansion of order 2κ0 of the power series u(σ (j,γ) ),
P 2κ 0
(j,γ)
(j,γ)
that is, ϕκ0 := m=0
u(xm )T m for 1 ≤ j ≤ Dγ .
According to Theorem 2.2, it follows that this
step requires performing
roughly O (nLγ + n3 )M(Dγ )M(Mγ )/ log(Mγ ) arithmetic operations in Q,
where Lγ denotes the number of arithmetic operations in Q required to evaluate the polynomials hi,γ of (5.16). Furthermore, in view of the application
of Lemma 2.4 it is important to remark that this step does not involve any
division by a nonconstant polynomial in the coefficients u1 , . . . , un .
(κ )
(κ )
Next we discuss Step II.3.b. Here we obtain approximations muγ0 , wuγ0,1 ,
(κ )
. . . , wuγ0,n of the polynomials that form the geometric solution of Vγ associated
with uγ = u(T γ1 X1 , . . . , T γn Xn ) with precision 2κ0 ≥ Mγ , namely
(κ )
(κ )
degY muγ0 = Dγ and degT muγ0 ≤ 2κ0 ,
(κ )
(κ )
degY wuγ0,i < Dγ and degT wuγ0,i ≤ 2κ0 for 1 ≤ i ≤ n,
(κ )
muγ0 ≡
(j,γ)
σi
QDγ
j=1
(κ )
(j,γ) Y − φ κ0
(j,γ) ≡ wuγ0,i T, φκ0
mod (T 2
mod (T 2
κ0 +1
κ0 +1
),
) for 1 ≤ i ≤ n.
159
(j,γ)
Here φκ0 is the Taylor expansion of φ(j,γ) := uγ (σ (j,γ) ) of order 2κ0 for
1 ≤ j ≤ Dγ .
From conditions (i)u,κ0– (iv)u,κ0 and elementary properties of the resultant
(κ )
it is easy to see that muγ0 satisfies the following identity:
0)
m(κ
uγ (Y
) = ResYe Y −
n
X
ui T
γi
i=1
(κ0 ) e
wu,i,γ
(Y ),
0) e
m(κ
u,γ (Y )
.
(5.43)
The resultant of the right-hand side is computed mod (T Mγ +1 ) by interpolation in the variable Y to reduce the problem to the computation of Dγ
resultants, as explained in the computation of the resultant in (5.34). These
Dγ resultants involve two polynomials of Q[T, Ye ] of degree in Ye bounded by
Dγ and are computed mod (T Mγ +1 ). Hence
we deduce that this step requires
roughly O M(Dγ )Dγ M(Mγ )/ log(Mγ ) arithmetic operations in Q.
We apply Lemma 2.4 in order to extend this procedure to an algorithm
(κ )
(κ )
(κ )
computing muγ0 , wuγ0,1 , . . . , wuγ0,n . For this purpose, we observe that a similar argument as in the proof of Proposition 5.9 proves that the denominators
in Q[Λ] which arise during the computation of the Dγ resultants
required
P
to compute the minimal polynomial of the generic version ni=1 Λi T γi Xi of
the linear form uγ are divisors of a polynomial of Q[Λ] of degree at most
4Dγ3 . Applying Theorem 2.1 we see that for a random choice of the coefficients u1 , . . . , un in the set {1, . . . , 4ρDγ3 } none of these denominators are
annihilated with probability at least 1 − 1/ρ.
Next, we consider Step II.3.c. For κ1 := ⌈log(2γn+1 E + 1)⌉, we apply
κ1 − κ0 times an adaptation of the global Newton iterator of [GLS01] to the
(κ )
(κ )
(κ )
polynomials muγ0 , wuγ0,1 , . . . , wuγ0,n computed in the previous step. In the kth
(k)
(k)
(k)
iteration step, we compute polynomials muγ , wuγ ,1 , . . . , wuγ ,n satisfying:
(k)
(k)
degY muγ = D and degT muγ ≤ 2k ,
(k)
m uγ =
QDγ
j=1 (Y
(j,γ)
− φk
(k)
),
(k)
degY wuγ ,i < D and degT wuγ ,1 ≤ 2k for 1 ≤ i ≤ n,
(j,γ)
σi
(k)
(j,γ)
≡ wuγ ,i (T, φk
) mod (T 2
k +1
) for 1 ≤ i ≤ n.
160
(j,γ)
Here φk is the Taylor expansion of φ(j,γ) := uγ (σ (j,γ) ) of order 2k for 1 ≤
(κ )
j ≤ Dγ . In particular, it follows that muγ1 is the required approximation to
muγ with precision 2γn+1 E.
Fix κ0 < k ≤ κ1 . We briefly describe how we can obtain an approximation
with precision 2k of the polynomials that form the geometric solution of Vγ
associated to the linear form uγ from an approximation with precision 2k−1 .
(k)
(k−1)
(k)
euγ ) − Y ,
Similarly to [GLS01], set ∆k (T, Y ) := uγ (w
euγ ) − uγ (wuγ ) = uγ (w
(k−1)
(k)
where w
euγ is the result of applying a “classical Newton step” to wuγ , as
k−1
(k)
described in [GLS01]. Furthermore, write ∆m (T, Y ) := T −1−2 (muγ −
k
(k)
(k−1)
(k−1)
muγ ). Since muγ (Y + ∆k ) ≡ 0 mod (T 2 +1 , muγ ) holds (see [DL08,
§4.2]), it follows that
(k−1)
0 ≡ m(k)
(Y + ∆k ) + T 2
uγ (Y + ∆k ) ≡ muγ
k−1 +1
mod (T 2
∆m (Y + ∆k )
k +1
, mu(k−1)
)
γ
(k−1)
≡ ∆k
∂muγ
k−1
(Y ) + T 2 +1 ∆m (Y )
∂Y
mod (T 2
k +1
, mu(k−1)
).
γ
We conclude that the following congruence relation holds:
∂m(k−1)
uγ
(k−1)
(k−1)
m(k)
≡
m
−
∆
mod
m
k
uγ
uγ
uγ
∂Y
mod (T 2
k +1
).
(5.44)
A similar argument proves the following congruence relation:
(k)
wuγ ,i
≡
(k−1)
w
euγ ,i
− ∆k
(k−1)
∂w
euγ ,i
∂Y
mod
mu(k−1)
γ
mod (T 2
k +1
) for 1 ≤ i ≤ n.
(5.45)
Each iteration of our adaptation of the global Newton iteration is based
on (5.44) and (5.45), which are extensions of the corresponding congruence
(k)
relations of [GLS01]. We first compute w
euγ by a standard Newton-Hensel
lifting, and then evaluate the expressions (5.44) and (5.45). With a similar
analysis as in [GLS01, Proposition 7] we conclude
that the whole procedure
3
requires roughly O (nLγ + n )M(Dγ )M(Eγn+1 ) arithmetic operations in Q.
Finally, combining the complexity estimates of Steps II.3.a, II.3.b, II.3.c
and the probability of achievement of the two generic conditions imposed
to the coefficients u1 , . . . , un (the condition underlying Lemma 5.16 and the
application of Lemma 2.4 in Step II.3.b), we deduce the statement of the
proposition.
161
Example. Consider the sparse polynomial system defined in (5.17) and their
associated inner normals γ (1) = (2, −1, 2), γ (2) = (−1, 2, 2) and γ (3) =
(−1, −1, 4). In (5.35) we have computed the geometric solutions for the
varieties V0,γ (i) (i = 1, 2, 3) associated to the linear form u := X1 − X2 .
From these geometric solutions, in the Step II of Algorithm 5.8 we obtain
approximations to the polynomials muγ (i) (i = 1, 2, 3). For the next step, to
compute a complete geometric solution of the variety associated to the linear
form u := X1 − X2 , we will deal with the first-order Taylor approximations
of the minimal polynomials of the generic linear form U := Λ1 X1 + Λ2 X2
centered at (Λ1 , Λ2 ) = (1, −1). Recall that, in this case, E = 3 (see (5.10)):
• For i = 1, we have γ (1) = (2, −1, 2), Dγ (1) = 2. Following (5.42), we
compute an approximation of T 2 mγ (1) (T 2 , T −1 Y ) with precision 12 by applying our modified Newton-Hensel lifting to the geometric solution of V0,γ (1)
previously computed, thus obtaining:
m1=Y 2+(−4T 11+4T 9+2T 3+ 4T 11+ 6T 9+2T 7+2T 3 )(Λ1 −1)+(8T 11+2T 9+2T 7 )(Λ2 +1) Y
−6T 12 + T 8 + T 4 − 1 + (−10T 12 − 6T 10 )(Λ1 − 1) + (2T 12 − 6T 10 − 2T 8 − 2T 4 + 2)(Λ2 + 1).
• For i = 2, we have γ (2) = (−1, 2, 2), Dγ (2) = 2. Following (5.42), we compute an approximation of T 2 mγ (2) (T 2 , T −1 Y ) with precision 12 by applying
our modified Newton-Hensel lifting to the geometric solution of V0,γ (2) , thus
obtaining:
m2 = Y 2+(4T 11−4T 9−2T 3+(8T 11+2T 9+2T 7 )(Λ1 −1) + (4T 11+6T 9+2T 7+2T 3 )(Λ2 +1))Y
−6T 12 + T 8 + T 4 − 1 + (−2T 12 + 6T 10 + 2T 8 + 2T 4 − 2)(Λ1 −1) + (10T 12 + 6T 10 )(Λ2 + 1).
• For i = 3, we have γ (3) = (−1, −1, 4), Dγ (3) = 4. Following (5.42), we
first compute an approximation of T 4 mγ (3) (T 4 , T −1 Y ) with precision 24 by
applying our modified Newton-Hensel lifting to the geometric solution of
V0,γ (3) , thus obtaining:
m3 := Y 4 + ((−12T 21 − 8T 17 − 4T 13 − 2T 5 )(Λ1 − 1) + (−12T 21 − 8T 17 − 4T 13 −
2T 5 )(Λ2 +1))Y 3 +(28T 22 −2T 14 +4T 10 −2T 6 +8T 2 +(−28T 22 +2T 14 −4T 10 +2T 6 −
8T 2 )(Λ2 + 1) + +(+28T 22 − 2T 14 + 4T 10 − 2T 6 + 8T 2 )(Λ1 − 1))Y 2 + ((−192T 23 −
70T 19 − 48T 15 − 2T 11 − 16T 7 − 8T 3 )(Λ1 − 1) + (−192T 23 − 70T 19 − 48T 15 − 2T 11 −
162
16T 7 − 8T 3 )(Λ2 + 1))Y + 152T 24 + 66T 20 − 32T 16 + 33T 12 − 16 − 8T 8 + (−304T 24 −
132T 20 + 64T 16 − 66T 12 + 16T 8 + 32)(Λ2 + 1) + (304T 24 + 132T 20 − 64T 16 + 66T 12 −
16T 8 − 32)(Λ1 − 1).
Using the algorithm of the statement of Proposition 5.18 for all γ ∈ Γ
we obtain approximations of the factors muγ which allow us to compute the
minimal polynomial m
b u and hence a geometric solution of Vb . Our next result
outlines this procedure, comprising Step III of Algorithm 5.8, and estimates
its complexity and error probability.
Proposition 5.19. Suppose that we are given a geometric solution of the
variety V0,γ for all γ ∈ Γ, as provided by Theorem 5.11, with a linear
form u ∈ Q[X1 , . . . , Xn ] whose coefficients are randomly chosen in the set
{1, . . . , 4ρD4 }, where ρ is a fixed positive integer. Then we can compute a
geometric solution of the curve Vb with
O (n3 N log Q + n4 )M(MΓ )M(D) M(D) + M(E)
arithmetic operations in Q and error probability bounded by 1/ρ. Here N :=
P
n
i=1 #∆i , Q := max1≤i≤n {kqk; q ∈ ∆i }, and MΓ := maxγ∈Γ max{γ1 , . . . ,γn+1 }.
Proof. For each γ ∈ Γ, we apply the algorithm underlying the proof of Proposition 5.18 in order to obtain an approximation of muγ with precision 2γn+1 E.
Due to Lemma 5.15, this polynomial immediately yields an approximation
with precision 2E of mγ (T, Y ) in Q((T 1/γn+1 ))[Y ].
Multiplying all these approximations,
Q we obtain an approximation with
precision 2E of the polynomial m
b u = γ∈Γ mγ of (5.40). Since every coefficient aj (T ) of m
b u ∈ Q(T )[Y ] is a rational function of Q(T ) having a reduced
representation with numerator and denominator of degree at most E, such
a representation of aj (T ) can be computed from its approximation with precision 2E using Padé approximation with O(M(E)) arithmetic operations in
Q.
In order to estimate the complexity of the whole procedure, we estimate
the complexity of its three main steps:
(i) the computation of the
P polynomials3 mγ with precision 2E for all γ ∈
Γ, which requires O
γ∈Γ (nLγ + n )M(Dγ ) M(Mγ )M(Dγ )/ log(Mγ ) +
M(Eγn+1 ) arithmetic operations in Q,
163
Q
(ii) the computation of the product γ∈Γ mγ with precision 2E, which
requires O M(D)M(E) arithmetic operations in Q,
(iii) the computation of a reduced representation
of all the coefficients of
m
b u ∈ Q(T )[Y ], which requires O M(E)D arithmetic operations in Q.
Observe that, from the sparse representation of the polynomials h1 , . . . , hn ,
we easily obtain a straight–line program computing the polynomials hi,γ of
(5.16) with O(nN
Pn log(QMγ )) arithmetic operations in Q for every γ ∈ Γ,
where N :=
i=1 #∆i and Q := max1≤i≤n {kqk; q ∈ ∆i }. Therefore,
the
2
3
algorithm performs O (n N log Q + n )M(MΓ )M(D) M(D) + M(E) arithmetic operations in Q, where MΓ := maxγ∈Γ {Mγ , γn+1 }.
Next we discuss how this procedure can be extended to the computation
of a geometric solution of Vb . Two computations of the above procedure
involve divisions by the coefficients ui of the linear form u: the computation
of the resultant of (5.43) for all γ ∈ Γ and the Padé approximations of
(iii). Both computations are reduced to D applications of the EEA, which
is performed in a ring Q(Λ). A similar analysis as in Proposition 5.9 shows
that all the denominators in Q[Λ] arising during such application of the
EEA are divisors of a polynomial of degree 4D4 . Therefore, according to
Lemma 2.4, we conclude that a geometric solution of Vb can be computed
3
4
with O (n N log Q+n )M(MΓ )M(D) M(D)+M(E) arithmetic operations
in Q, with an algorithm with error probability at most 1/ρ, provided that
the coefficients of u are randomly chosen in the set {1, . . . , 4ρD4 }.
Example. We continue with our previous example.
• For i = 1, the algorithm obtains an approximation m
e γ (1) of mγ (1) by substituting Y = T Y in the polynomial m1 previously computed, multiplying it
by T −2 and replacing T 2 with T , which yields:
m
e γ (1) =Y 2+(−4T 5+4T 4+2T +(8T 5+2T 4+2T 3 )(Λ2 +1)+(4T 5+6T 4+2T 3+2T )(Λ1 −1))Y
2
1
−6T 5 + T 3 + T − + (2T 5 − 6T 4 − 2T 3 − 2T + )(Λ2 + 1) + (−10T 5 − 6T 4 )(Λ1 − 1).
T
T
e γ (2) of mγ (2) by substituting Y = T Y in the polynomial m2 , multiplying it by T −2 and replacing
T 2 with T , which yields:
164
m
e γ (2) = Y 2+(4T 5−4T 4−2T + (8T 5+2T 4+2T 3 )(Λ1 −1) + (4T 5+6T 4+2T 3+2T )(Λ2 + 1))Y
1
2
−6T 5 + T 3 + T − + (−2T 5 + 6T 4 + 2T 3 + 2T − )(Λ1 − 1) + (10T 5 + 6T 4 )(Λ2 + 1).
T
T
e γ (3) of mγ (3) by substituting Y = T Y in the polynomial m3 , multiplying it by T −4 and replacing
T 4 with T , which yields:
m
e γ (3) = Y 4 + ((−12T 5− 8T 4− 4T 3−2T )(Λ1 − 1) + (−12T 5−8T 4−4T 3 −2T )(Λ2 + 1))Y 3 +
(28T 5−2T 3+4T 2−2T+8+(28T 5−2T 3+4T 2−2T +8)(Λ1−1)+(−28T 5−2T 3−4T 2+2T−8)(Λ2+1))Y 2
+((−192T 5−70T 4−48T 3−2T 2−16T−8)(Λ1 −1)+(−192T 5−70T 4−48T 3−2T 2−16T−8)(Λ2 +1))Y
16
32
+152T 5+ 66T 4− 32T 3+ 33T 2− 8T− + (304T 5+ 132T 4− 64T 3+ 66T 2− 16T− )(Λ1 − 1)
T
T
32
5
4
3
2
+(−304T − 132T + 64T − 66T + 16T + )(Λ2 + 1).
T
Computing the first-order Taylor approximation centered at (Λ1 , Λ2 ) =
(1, −1) of the product m
e γ (1) m
e γ (2) m
e γ (3) with precision 2E = 6 in the variable
T , and applying a Padé approximation algorithm, we obtain the polynomial
8T − 2 6 2T 2 − 32T + 1 4 −28T 2 − 2T + 40 2 33T 3 + 24T 2 − 16
M := Y 8 +
Y +
Y +
Y +
+
T
T2
T2
T3
8T−2
2
2
2
4T −64T+2 4 −48T+14 3 −84T −6T+120 2 14T +80T−8
Y 6−10Y 5+
Y +
Y +
Y +
Y+
T
T2
T
T2
T2
2
3
2
−8T + 2 6
−4T + 64T − 2 4 −48T + 14 3
132T + 96T − 64
(Λ1 −1)+(
Y +
Y −10Y 5 +
Y +
T3
T
T2
T
84T 2 + 6T − 120 2 14T 2 + 80T − 8
−132T 3 − 96T 2 + 64
Y +
Y +
)(Λ2 + 1).
2
2
T
T
T3
This polynomial is the first-order Taylor approximation centered at (Λ1 , Λ2 ) =
(1, −1) of the minimal polynomial of the generic linear form U := Λ1 X1 +
Λ2 X2 .
Therefore, a geometric solution of the curve Vb defined in (5.19) is given
by the polynomials
where
m
b u (Y ),
∂m
bu
X1 + vb1 (Y ),
∂Y
∂m
bu
X2 + vb2 (Y ),
∂Y
165
2
2
2
+1 4
+40 2
m
b u = Y 8 + 8TT−2 Y 6 + 2T −32T
Y + −28T T−2T
Y + 24T +33T
2
T2
T3
the polynomial obtained substituting Λ1 = 1, Λ2 = −1 in M ,
3 −16
2
2
+2 4
+120 2
8T −2 6
Y − 10Y 5 + 4T −64T
Y + −48TT +14 Y 3 + −84T −6T
Y
T
T2
T2
3
2
2
132T +96T −64
14T +80T −8
Y +
is the partial derivative ∂M /∂Λ1 ,
T2
T3
vb1 =
2
2
−8T +2 6
−2 4
−120 2
Y − 10Y 5 + −4T +64T
Y + −48TT +14 Y 3 + 84T +6T
Y
T
T2
T2
3
2
2
−132T −96T +64
14T +80T −8
Y +
is the partial derivative ∂M /∂Λ2 .
T2
T3
vb2 =
is
+
+
Putting together Theorem 5.11 and Proposition 5.19 we obtain the main
result of this section.
Theorem 5.20. Let ρ be a fixed positive integer. Suppose that the coefficients of the linear form u
e of the statement of Theorem 5.11 and of the linear
form u are randomly chosen in the set {1, . . . , 4nρD4 }. Then the algorithm
underlying Theorem 5.11 and Proposition 5.19 computes a geometric solution of the curve Vb with error probability 3/ρ performing O (n3 N log Q +
4
n
arithmetic operations in Q. Here N :=
Γ )M(D) M(D) + M(E)
P)M(M
n
i=1 #∆i , Q := max1≤i≤n {kqk; q ∈ ∆i }, and MΓ := maxγ∈Γ kγk.
5.2.6.
Solving the generic sparse system
Now we obtain a geometric solution of the zero-dimensional variety V1 :=
{x ∈ Cn : h1 (x) = 0, . . . , hn (x) = 0} from a geometric solution of the curve
Vb .
With notations as in the previous section, we have that V1 = π −1 (1),
where π : Vb → A1 is the linear projection defined by π(x, t) := t. Moreover,
due to Lemma 5.13, the equality V1 = π −1 (1) ∩ V (I) holds.
This enables us to easily obtain a geometric solution of V1 from a geometric solution of the curve Vb . Indeed, let m
b u (T, Y ), vb1 (T, Y ), . . . , vbn (T, Y )
be the polynomials which form a geometric solution of Vb associated to a
linear form u ∈ Q[X]. Suppose further that the linear form u separates
the points of V1 . Making the substitution T = 1, we obtain new polynomials m
b u (1, Y ), vb1 (1, Y ), . . . , vbn (1, Y ) ∈ Q[Y ] such that m
b u (1, u(X)) and
166
∂m
bu
(1, u(X))Xi
∂Y
− vbi (1, u(X)) (1 ≤ i ≤ n) vanish over V1 . Taking into account that degY (m
b u ) = D = #V1 and that u separates the points of V1 , it
follows that the polynomials m
b u (1, Y ), vb1 (1, Y ), . . . , vbn (1, Y ) ∈ Q[Y ] form a
geometric solution of V1 .
Proposition 5.21. Let ρ be a fixed positive integer. With assumptions and
notations as in Theorem 5.20, the algorithm described above computes a geometric solution of the zero-dimensional variety V1 with
error probability 4/ρ
3
4
using O (n N log Q+n )M(MΓ )M(D) M(D)+M(E) arithmetic operations
in Q.
Example. By substituting 1 for T in the geometric solution of the curve Vb
defined in (5.19) computed in the previous section, we obtain a geometric
solution of the zero-dimensional variety V1 = {(x1 , x2 ) ∈ C2 : 1 − x21 − x22 −
x21 x22 = 0, 1 + x21 x2 + x1 x22 = 0} defined by the system (5.17), namely,
mu (Y ),
∂mu
(Y )X1 + v1 (Y ),
∂Y
∂mu
(Y )X2 + v2 (Y ),
∂Y
where
mu (Y ) := m
b u (1, Y ) = Y 8 + 6Y 6 − 29Y 4 + 10Y 2 + 41,
v1 (Y ) := vb1 (1, Y ) = 6Y 6 − 10Y 5 − 58Y 4 − 34Y 3 + 30Y 2 + 86Y + 164,
5.3.
v2 (Y ) := vb2 (1, Y ) = −6Y 6 − 10Y 5 + 58Y 4 − 34Y 3 − 30Y 2 + 86Y − 164.
The solution of the input system
Let notations and assumptions be as in the previous sections. Assume
that we are given a geometric solution mu (Y ), v1 (Y ), . . . , vn (Y ) of the zerodimensional variety V1 defined by the polynomials h1 = f1 + g1 , . . . , hn =
fn + gn . Assume further that the linear form u of such a geometric solution
separates the points of the zero-dimensional variety f1 = 0, . . . , fn = 0. In
this section we describe a procedure for computing a geometric solution of
the input system f1 = 0, . . . , fn = 0.
For this purpose, we introduce an indeterminate T over Q[X] and consider
the “deformation” F1 , . . . , Fn ∈ Q[X, T ] of the polynomials f1 , . . . , fn defined
5.3. THE SOLUTION OF THE INPUT SYSTEM
167
in the following way:
Fi (X, T ) := fi (X) + (1 − T )gi (X) (1 ≤ i ≤ n).
(5.46)
Set V := {(x, t) ∈ An+1 : F1 (x, t) = 0, . . . , Fn (x, t) = 0} and denote by π :
V → A1 the projection map defined by π(x, t) := t. As in Subsection 5.2.5, we
introduce the variety Vdom ⊂ An+1 defined as the union of all the irreducible
components of V whose projection over A1 is dominant.
5.3.1.
Solution of the second deformation
In this section we describe an efficient procedure for computing a geometric solution of Vdom from the geometric solution of π −1 (0) provided by
Proposition 5.21.
Since π −1 (0) is the variety defined by the “sufficiently generic” sparse
system h1 (X) = F1 (X, 0) = 0, . . . , hn (X) = Fn (X, 0) = 0, with similar
arguments to those leading to the proof of Lemma 5.13, it is not difficult to
see that the polynomials F1 , . . . , Fn , the variety V, the projection π : V → A1 ,
and the fiber π −1 (0) satisfy all the assumptions of Lemma 5.12. We conclude
that Vdom is a curve and that the identity V ∩ π −1 (0) = Vdom ∩ π −1 (0) holds.
Furthermore, Lemma 5.12 implies that all the hypotheses of [Sch03, Theorem
2] are satisfied.
Therefore, applying the “formal Newton lifting process” underlying Theorem 2.2, we compute polynomials m
e u (T, Y ), ve1 (T, Y ), . . . , ven (T, Y ) ∈ Q[T, Y ]
which form a geometric solution of Vdom . The formal Newton lifting process
requires O (nL′ + n4 )M(D)M(E ′ ) arithmetic operations in Q, where L′ denotes the number of arithmetic operations required to evaluate F1 , . . . , Fn
and E ′ is any upper bound of the degree of m
e u in the variable T .
In order to estimate the quantity L′ , we observe that from the sparse
representation of the polynomials f1 , . . . , fn , h1 , . . . , hn we easily obtain a
straight–line program of length at most O(nN log Q) which evaluates f1 , . . . ,
fn , h1 , . . . , hn . Therefore, the polynomials F1 , . . . , Fn can also be represented
by a straight–line program of length at most O(nN log Q).
Furthermore, we can apply Lemma 5.3 in order to estimate degT m
e u in
e1 , . . . , Q
en ⊂ Rn+1 be the Newton polytopes
combinatorial terms. Indeed, let Q
of F1 , . . . , Fn and let ∆ ⊂ Rn+1 be the standard n–dimensional simplex in
168
ei ⊂ Qi × [0, 1] holds for 1 ≤ i ≤ n, where
the hyperplane {T = 0}. Since Q
n
Qi ⊂ R is the Newton polytope of hi , by (5.5) of Lemma 5.3 we deduce the
following estimate:
′
degT m
e u ≤ E :=
n
X
M(∆, Q1 , . . . , Qi−1 , Qi+1 , . . . , Qn ).
(5.47)
i=1
With this estimate for L′ and this definition of E ′ , we have the following
result.
Proposition 5.22. Suppose that we are given a geometric solution of the variety V1 , as provided by Proposition 5.21. A geometric solution ofVdom can be
deterministically computed with O (n2 N log Q + n4 )M(D)M(E ′ ) arithmetic
operations in Q.
5.3.2.
Solving the input system
Making the substitution T = 1 in the polynomials m
e u (T, Y ), vei (T, Y ) (1 ≤
i ≤ n) which form the geometric solution of Vdom computed by the algorithm
of Proposition 5.22 we obtain polynomials m
e u (1, Y ), ve1 (1, Y ), . . . , ven (1, Y ) ∈
Q[Y ] which represent a complete description of our input system f1 (X) =
0, . . . , fn (X) = 0, eventually including multiplicities. Such multiplicities
are represented by multiple factors of m
e u (1, Y ), which are also factors of
ve1 (1, Y ), . . . , ven (1, Y ) (see, e.g., [GLS01, §6.5]). In order to remove them,
we compute a(Y ) := gcd m
e u (1, Y ), (∂ m
e u /∂Y )(1, Y ) , and the polynomials
m(Y ) := m
e u (1, Y )/a(Y ), b(Y ) :=
e u /∂Y )(1, Y )/a(Y ) −1 mod m(Y ),
(∂ m
and wi (Y ) := b(Y ) vei (1, Y )/a(Y ) mod m(Y ) (1 ≤ i ≤ n). Then m, w1 , . . . ,
wn form a geometric
solution of our input system and can be computed with
′
O nM(D)E additional arithmetic operations in Q.
Summarizing, we sketch the whole procedure computing a geometric solution of the input system f1 = 0, . . . , fn = 0. Fix ρ ≥ 4. We randomly choose
the coefficients of the polynomials g1 , . . . , gn in the set {1, . . . , 4ρ(nd)2n+1 +
2ρn2 2N1 +···+Ns } and coefficients of linear forms u, u
e in the set {1, . . . , 16nρD4 }.
By Theorem 2.1 it follows that the polynomials g1 , . . . , gn and the linear forms
u, u
e satisfy all the conditions required with probability at least 1 − 1/ρ. Then
we apply the algorithms underlying Propositions 5.21 and 5.22 in order to
obtain a geometric solution of the variety Vdom . Finally, we use the procedure
5.3. THE SOLUTION OF THE INPUT SYSTEM
169
above to compute a geometric solution of the input system f1 = 0, . . . , fn = 0.
This yields the following result.
Theorem 5.23. The algorithm sketched above computes a geometric solution
of the input system f1 = 0, . . . , fn = 0 with error probability at most 1/ρ using
O n3 N log Q + n4 M(D) M(MΓ ) M(D) + M(E) + M(E ′ )
Pn
arithmetic operations in Q. Here N :=
i=1 #∆i , MΓ := maxγ∈Γ kγk,
Q := max1≤i≤n {kqk; q ∈ ∆i } and E, E ′ are defined in (5.36) and (5.47)
respectively.
We remark that our algorithm can be applied mutatis mutandis in order to compute the isolated points of an input system having a solution
set with positive-dimensional components. Indeed, since the first deformation is not determined by the input system but by its monomial structure, it computes a geometric solution of a generic sparse system as described in Section 5.2. Then we execute our second deformation on the
polynomials F1 , . . . , Fn of (5.46), considering the saturation (I : J ∞ ), where
I := (F1 , . . . , Fn ) ⊂ Q[X, T ] and J denotes the Jacobian determinant of
F1 , . . . , Fn with respect to the variables X. From Lemma 5.12 it follows that
positive–dimensional components of f1 = 0, , . . . , fn = 0 are “cleaned” by
the saturation (I : J ∞ ). Hence, our algorithm properly outputs the isolated
points of f1 = 0, , . . . , fn = 0, as stated.
170
Bibliography
[ABRW96]
M.E. Alonso, E. Becker, M.-F. Roy, and T. Wörmann. Zeroes,
multiplicities and idempotents for zerodimensional systems. In
Algorithms in Algebraic Geometry and Applications, Proceedings of MEGA’94, volume 143 of Progr. Math., pages 1–15,
Boston, 1996. Birkhäuser.
[ANMR91]
M.E. Alonso, G. Niesi, T. Mora, and M. Raimondo. Local
parametrizations of space curves at singular points. In Computer Graphics and Mathematics, Eurographic Seminar Series,
pages 61–90, Berlin Heidelberg New York, 1991. Springer.
[ANMR92]
M.E. Alonso, G. Niesi, T. Mora, and M. Raimondo. An algorithm for computing analytic branches of space curves at singular points. In Proceedings 1992 International Workshop on
Mathematics Mechanization, Beijing, China, July 16-18, 1992,
pages 135–166. International Academic Publishers, 1992.
[Art76]
M. Artin. Lectures on deformation of singularities, volume 54
of Tata Inst. Fundam. Res. Lect. Math. Tata Inst. Fund. Res.,
Bombay, 1976.
[Bar04]
M. Bardet. Etude des systèmes algébriques surdéterminés. Applications aux codes correcteurs et à la cryptographie. PhD thesis, Université Paris 6, 2004.
[BCS97]
P. Bürgisser, M. Clausen, and M.A. Shokrollahi. Algebraic
Complexity Theory, volume 315 of Grundlehren Math. Wiss.
Springer, Berlin, 1997.
171
172
BIBLIOGRAPHY
[BCSS98]
L. Blum, F. Cucker, M. Shub, and S. Smale. Complexity and
Real Computation. Springer, New York Berlin Heidelberg, 1998.
[BDG88]
J.L. Balcázar, J. Díaz, and J. Gabarró. Structural complexity
I, volume 11 of Monogr. Theoret. Comput. Sci. EATCS Ser.
Springer, Berlin, 1988.
[Ber75]
D.N. Bernstein. The number of roots of a system of equations.
Funct. Anal. Appl., 9:183–185, 1975.
[BLS03]
A. Bostan, G. Lecerf, and E. Schost. Tellegen’s principle into
practice. In J.R. Sendra, editor, Proceedings of the International Symposium on Symbolic and Algebraic Computation (ISSAC’03) (Philadelphia, USA, August 3–6, 2003), pages 37–44,
New York, 2003. ACM Press.
[BM93]
D. Bayer and D. Mumford. What can be computed in algebraic
geometry? In D. Eisenbud and L. Robbiano, editors, Computational Algebraic Geometry and Commutative Algebra, volume
XXXIV of Sympos. Math., pages 1–49, Cambridge, 1993. Cambridge Univ. Press.
[BM96]
D. Bini and B. Mourrain.
Polynomial
http://www-sop.inria.fr/saga/POL, 1996.
test
suite.
[BMWW04] A. Bompadre, Guillermo Matera, Rosita Wachenchauzer, and
Ariel Waissbein. Polynomial equation solving by lifting procedures for ramified fibers. Theor. Comput. Sci., 315(2-3):335–
369, 2004.
[Bom00]
A. Bompadre. Un problema de eliminación geométrica en sistemas de Pham–Brieskorn. Master’s thesis, Facultad de Ciencias Exactas y Naturales, Universidad de Buenos Aires, 2000.
[BP94]
D. Bini and V. Pan. Polynomial and matrix computations.
Progress in Theoretical Computer Science. Birkhäuser, Boston,
1994.
[BP06]
C. Beltrán and L.M. Pardo. Non universal polynomial equation
solving. In L.M. Pardo et al., editor, Foundations of Computational Mathematics, Santander 2005, volume 331 of London
BIBLIOGRAPHY
173
Math. Soc. Lecture Note Ser., pages 1–35, Cambridge, 2006.
Cambridge Univ. Press.
[BR01]
J. Fernandez Bonder and J. Rossi. Blow-up vs. spurious steady
solutions. Proc. Amer. Math. Soc., 129(1):139—144, 2001.
[Bro87]
D.W. Brownawell. Bounds for the degree in the Nullstellensatz.
Ann. of Math., 126:577–591, 1987.
[BS83]
W. Baur and V. Strassen. The complexity of partial derivatives.
Theoret. Comput. Sci., 22:317–330, 1983.
[BS05]
A. Bostan and E. Schost. Polynomial evaluation and interpolation on special sets of points. J. Complexity, 21(4):420–446,
2005.
[Buc85]
B. Buchberger. Gröbner bases: An algoritmic method in polynomial ideal theory. In N. K. Bose et al, editor, Multidimensional System Theory, pages 374–383. Reidel, Dordrecht, 1985.
[BW93]
T. Becker and V. Weispfenning. Gröbner bases. A computational approach to commutative algebra, volume 141 of Grad.
Texts in Math. Springer, New York, 1993.
[Can84]
J. Cannon. The one-dimensional heat equation. Cambridge
University Press, Cambridge, 1984.
[CDS96]
E. Cattani, A. Dickenstein, and B. Sturmfels. Computing multidimensional residues. In T. Recio and L. González-Vega, editors, Algorithms in Algebraic Geometry and Applications, Proceedings of the MEGA-94 conference, Santander, Spain, April
5-9, 1994, volume 143 of Progr. Math., Basel, 1996. Birkhäuser.
[CFQ91a]
M Chipot, M Fila, and P Quittner. Stationary solutions.
blow up and convergence to stationary solutions for semilinear
parabolic equations with nonlinear boundary conditions. Acta
Math Univ Comenian, 60(1):35–103, 1991.
[CFQ91b]
M. Chipot, M. Fila, and P. Quittner. Stationary solutions,
blow up and convergence to stationary solutions for semilinear
parabolic equations with nonlinear boundary conditions. Acta
Mathematica Universitatis Comenianae, 60(1):35–103, 1991.
174
BIBLIOGRAPHY
[CGH91]
L. Caniglia, A. Galligo, and J. Heintz. Equations for the projective closure and effective Nullstellensatz. Discrete Appl. Math.,
33:11–23, 1991.
[CGH+ 03]
D. Castro, M. Giusti, J. Heintz, G. Matera, and L.M. Pardo.
The hardness of polynomial equation solving. Found. Comput.
Math., 3(4):347–420, 2003.
[CHMP01]
D. Castro, K. Hägele, J.E. Morais, and L.M. Pardo. Kronecker’s
and Newton’s approaches to solving: a first comparison. J.
Complexity, 17(1):212–303, 2001.
[CJPS02]
D. Castro, J.L. Montaña, L.M. Pardo, and J. San Martín.
The distribution of condition numbers of rational data of
bounded bit length. Foundations of Computational Mathematics, 2(1):1—52, 2002.
[CLO98]
D. Cox, J. Little, and D. O’Shea. Using algebraic geometry,
volume 185 of Grad. Texts in Math. Springer, New York, 1998.
[CS99]
F. Cucker and S. Smale. Complexity estimates depending on
condition and round–off error. J. ACM, 46(1):113–184, 1999.
[Dan94]
V. Danilov. Algebraic varieties and schemes. In I.R. Shafarevich, editor, Algebraic Geometry I, volume 23 of Encyclopaedia
of Mathematical Sciences, pages 167–307. Springer, Berlin Heidelberg New York, 1994.
[Ded97]
J.-P. Dedieu. Condition number analysis for sparse polynomial
systems. In F. Cucker and M. Shub, editors, Foundations of
computational mathematics (Rio de Janeiro, 1997), pages 267–
276, Berlin Heidelberg New York, 1997. Springer.
[DL08]
C. Durvye and G. Lecerf. A concise proof of the Kronecker
polynomial system solver from scratch. Expo. Math., 26(2):101–
139, 2008.
[DM09]
Ezequiel Dratman and Guillermo Matera. On the solution of
the polynomial systems arising in the discretization of certain
odes. Computing, 85(4):301–337, July 2009.
BIBLIOGRAPHY
175
[DMW09]
E. Dratman, G. Matera, and A. Waissbein. Robust algorithms
for generalized Pham systems. Comput. Complexity, 18(1):105–
154, 2009.
[Dra10]
Ezequiel Dratman. Approximation of the solution of certain
nonlinear {ODEs} with linear complexity. Journal of Computational and Applied Mathematics, 233(9):2339 – 2350, 2010.
[Dra13a]
Ezequiel Dratman. Efficient approximation of the solution of
certain nonlinear reaction–diffusion equations with small absorption. Journal of Complexity, 29(3–4):263 – 282, 2013.
[Dra13b]
Ezequiel Dratman. Efficient approximation of the solution of
certain nonlinear reaction-diffusion equations with large absorption. J. Comput. Appl. Math., 238:180–202, January 2013.
[Duv89]
D. Duval. Rational Puiseux expansions.
70:119–154, 1989.
[Duv90]
J. Duvallet. Computation of solutions of two–point boundary
value problems by a simplicial homotopy algorithm. In K. Georg
E. Allgower, editor, Computational Solution of Nonlinear Systems of Equations, volume 26 of Lectures Appl. Math., pages
135–150, Providence, RI., 1990. Amer. Math. Soc.
[EH99]
D. Eisenbud and J. Harris. The Geometry of Schemes, volume
197 of Grad. Texts in Math. Springer, New York, 1999.
[Eis95]
D. Eisenbud. Commutative Algebra with a View Toward Algebraic Geometry, volume 150 of Grad. Texts in Math. Springer,
New York, 1995.
[Ewa96]
G. Ewald. Combinatorial convexity and algebraic geometry, volume 168 of Grad. Texts in Math. Springer, New York, 1996.
[FGR02]
R. Ferreira, P. Groisman, and J.D. Rossi. Numerical blow–
up for a nonlinear problem with a nonlinear boundary condition. Mathematical models and methods in applied sciences,
12(4):461–484, 2002.
Compos. Math.,
176
BIBLIOGRAPHY
[Ful84]
W. Fulton. Intersection Theory. Springer, Berlin Heidelberg
New York, 1984.
[GH01]
M. Giusti and J. Heintz. Kronecker’s smart, little black–boxes.
In A. Iserles R. Devore and E. Süli, editors, Proceedings of Foundations of Computational Mathematics, FoCM’99, Oxford 1999,
volume 284 of London Math. Soc. Lecture Note Ser., pages 69–
104, Cambridge, 2001. Cambridge Univ. Press.
[GHH+ 97]
M. Giusti, K. Hägele, J. Heintz, J.E. Morais, J.L. Montaña, and
L.M. Pardo. Lower bounds for Diophantine approximation. J.
Pure Appl. Algebra, 117,118:277–317, 1997.
[GHM+ 98]
M. Giusti, J. Heintz, J.E. Morais, J. Morgenstern, and L.M.
Pardo. Straight–line programs in geometric elimination theory.
J. Pure Appl. Algebra, 124:101–146, 1998.
[GHMP97]
M. Giusti, J. Heintz, J.E. Morais, and L.M. Pardo. Le rôle des
structures de données dans les problèmes d’élimination. C. R.
Math. Acad. Sci. Paris, 325:1223–1228, 1997.
[Giu89]
M. Giusti. Complexity of standard bases in projective dimension
zero. In J.H. Davenport, editor, Proceedings of the European
Conference on Computer Algebra, volume 378 of Lecture Notes
in Comput. Sci., pages 333–335, Berlin, 1989. Springer.
[Giu91]
M. Giusti. Complexity of standard bases in projective dimension
zero II. In S. Sakata, editor, Proceedings of Applied Algebra,
Algebraic Algorithms and Error–Correcting Codes, AAECC–8,
Aug 20–24, 1990, Tokyo, Japan, volume 508 of Lecture Notes
in Comput. Sci., pages 322–328, Berlin, 1991. Springer.
[GKZ94]
I.M. Gelfand, M.M. Kapranov, and A.V. Zelevinsky. Discriminants, Resultants, and Multidimensional Determinants.
Birkhäuser, Boston, 1994.
[GLGV98]
M.-J. Gonzalez-Lopez and L. Gonzalez-Vega. Newton identities
in the multivariate case: Pham Systems. In B. Buchberger et al.,
editor, Gröbner Bases and Applications, volume 251 of London
Math. Soc. Lecture Note Ser., pages 351–366. Cambridge Univ.
Press, Cambridge, 1998.
BIBLIOGRAPHY
177
[GLS01]
M. Giusti, G. Lecerf, and B. Salvy. A Gröbner free alternative
for polynomial system solving. J. Complexity, 17(1):154–211,
2001.
[GP74]
V. Guillemin and A. Pollack. Differential Topology. Prentice–
Hall, Englewood Cliffs, NJ, 1974.
[GV98]
L. Gonzalez-Vega. A special quantifier elimination algorithm
for Pham systems. In C. Detzell et al., editor, Real algebraic
geometry and ordered structures, AMS special session, Baton
Rouge, LA, USA, April 17-21, volume 253 of Contemp. Math.,
pages 115–366, Providence, RI, 1998. Amer. Math. Soc.
[Hei83]
J. Heintz. Definability and fast quantifier elimination in algebraically closed fields. Theoret. Comput. Sci., 24(3):239–277,
1983.
[Hen81]
D. Henry. Geometric theory of semilinear parabolic equations,
volume 840 of Lecture Notes in Mathematics. Springer, Berlin
Heideilberg, 1981.
[HJS+ 02]
J. Heintz, G. Jerónimo, J. Sabia, J. San Martín, and P. Solernó. Intersection theory and deformation algorithms. the multihomogeneous case. Manuscript Universidad de Buenos Aires,
2002.
[HJS13]
MaríA Isabel Herrero, Gabriela Jeronimo, and Juan Sabia.
Affine solution sets of sparse polynomial systems. J. Symb.
Comput., 51:34–54, April 2013.
[HKP+ 00]
J. Heintz, T. Krick, S. Puddu, J. Sabia, and A. Waissbein. Deformation techniques for efficient polynomial equation solving.
J. Complexity, 16(1):70–109, 2000.
[HKR11]
J. Heintz, B. Kuijpers, and A. Rojas Paredes. Software engineering and complexity in effective algebraic geometry. Computer
Research Repository abs/1110.3030, 2011.
[HKR12]
J. Heintz, B. Kuijpers, and A. Rojas Paredes. On the intrinsic
complexity of elimination problems in effective algebraic geometry. Computer Research Repository abs/1201.4344, 2012.
178
BIBLIOGRAPHY
[HM93]
J. Heintz and J. Morgenstern. On the intrinsic complexity of
elimination theory. J. Complexity, 9:471–498, 1993.
[HM00]
T. Sauer H.M. Möller. H-bases for polynomial interpolation and
system solving. Adv. Comput. Math., 12(4):335–362, 2000.
[HMPW98] J. Heintz, G. Matera, L.M. Pardo, and R. Wachenchauzer. The
intrinsic complexity of parametric elimination methods. Electron. J. SADIO, 1(1):37–51, 1998.
[HMW01]
J. Heintz, G. Matera, and A. Waissbein. On the time–space
complexity of geometric elimination procedures. Appl. Algebra
Engrg. Comm. Comput., 11(4):239–296, 2001.
[HS95]
B. Huber and B. Sturmfels. A polyhedral method for solving
sparse polynomial systems. Math. Comp., 64(112):1541–1555,
1995.
[HS97]
B. Huber and B. Sturmfels. Bernstein’s Theorem in affine space.
Discrete Comput. Geom., 17:137–141, 1997.
[JMSW09]
G. Jeronimo, G. Matera, P. Solernó, and A. Waissbein. Deformation techniques for sparse systems. Found. Comput. Math,
9:1–50, 2009.
[Kho78]
A.G. Khovanski. Newton polyhedra and the genus of complete
intersections. Funct. Anal. Appl., 12:38–46, 1978.
[KM96]
K. Kühnle and E. Mayr. Exponential space computation of
Gröbner bases. In Proceedings of the 1996 International Symposium on Symbolic and Algebraic Computation, ISSAC’96
(Zürich, 24–26 July 1996), volume 358 of ACM Press, pages
63–71, New York, 1996. ACM Press.
[KP96]
T. Krick and L.M. Pardo. A computational method for Diophantine approximation. In L. González-Vega and T. Recio,
editors, Algorithms in Algebraic Geometry and Applications,
Proceedings of MEGA’94, volume 143 of Progr. Math., pages
193–254, Boston, 1996. Birkhäuser Boston.
BIBLIOGRAPHY
179
[Küh98]
K. Kühnle. Space optimal computation of normal forms of
polynomials. PhD thesis, Technischen Universität München,
München, Germany, 1998.
[Kus76]
A.G. Kushnirenko. Newton polytopes and the Bézout Theorem.
Funct. Anal. Appl., 10:233–235, 1976.
[Laz83]
D. Lazard. Gröbner bases, Gaussian elimination and resolution
of systems of algebraic equations. In Computer algebra (London,
1983), number 162 in Lecture Notes in Comput. Sci., pages 146–
156, Berlin, 1983. Springer.
[Lec02]
G. Lecerf. Quadratic Newton iteration for systems with multiplicity. Found. Comput. Math., 2(3):247–293, 2002.
[Lec03]
G. Lecerf. Computing the equidimensional decomposition of an
algebraic closed set by means of lifting fibers. J. Complexity,
19(4):564–596, 2003.
[LL01]
T.-Y. Li and X. Li. Finding mixed cells in the mixed volume
computation. Found. Comput. Math., 1(2):161–181, 2001.
[LV98]
M. J. Gonzalez Lopez and L. Gonzalez Vega. Newton identities
in the multivariate case: Pham systems. London Mathematical
Society Lecture Notes Series, 251:351–366, 1998.
[LW96]
T.Y. Li and X. Wang. The BKK root count in Cn . Math.
Comp., 65(216), 1996.
[Mat80]
H. Matsumura. Commutative Algebra. Benjamin, 1980.
[Mat86]
H. Matsumura. Commutative Ring Theory. Cambridge Univ.
Press, Cambridge, 1986.
[May89]
E. Mayr. Membership in polynomial ideals over Q is Exponential Space Complete. In B. Monien et al., editor, Proceedings of
the 6th Annual Symposium on Theoretical Aspects of Computer
Science (STACS’89), Paderborn (FRG) 1989, number 349 in
Lecture Notes in Comput. Sci., pages 400–406, Berlin, 1989.
Springer.
180
BIBLIOGRAPHY
[MD04]
G. Matera M. De Leo, E. Dratman. On the numerical solution of certain nonlinear systems arising in semilinear parabolic
PDEs. In Anales JAIIO (Jornadas Argentinas de Informática
e Investigación Operativa, 2004.
[MD05]
G. Matera M. De Leo, E. Dratman. Numeric vs. symbolic homotopy algorithms in polynomial system solving: A case study.
J. Complexity, 21(4):502–531, 2005.
[Miz08]
Tomohiki Mizutani. Finding All Mixed Cells for Polyhedral Homotopies. PhD thesis, Tokyo Institute of Technology, 2008.
[MM82]
E. Mayr and A. Meyer. The complexity of the word problem
for commutative semigroups. Adv. Math., 46:305–329, 1982.
[MP97]
B. Mourrain and V. Pan. Solving special polynomial systems by
using structural matrices and algebraic residues. In F. Cucker
and M. Shub, editors, Proceedings Foundations of Computational Mathematics (FOCM’97), pages 287–304, Berlin Heidelberg New York, 1997. Springer.
[MP00]
B. Mourrain and V. Pan. Multivariate polynomials, duality and
structured matrices. J. Complexity, 16(1):110–180, 2000.
[MPT92]
T. Mora, G. Pfister, and C. Traverso. An introduction to the
tangent cone algorithm. In C. Hoffmann, editor, Issues in
robotics and non-linear geometry, volume 6 of Advances in Computing Research, pages 199–270. JAI Press, Greenwich Conn.,
1992.
[MR04]
G. Malajovich and J.M. Rojas. High probability analysis of
the condition number of sparse polynomial systems. Theor.
Comput. Sci., 315(2–3):525–555, 2004.
[MT00]
B. Mourrain and P. Trebuchet. Solving projective complete
intersection faster. In Proceedings 2000 ACM-SIGSAM International Symposium on Symbolic and Algebraic Computation
ISSAC’2000 (August 6 - 10, 2000, St. Andrews, United Kingdom), pages 234–241, New York, 2000. ACM Press.
BIBLIOGRAPHY
181
[MTK07]
Tomohiko Mizutani, Akiko Takeda, and Masakazu Kojima. Dynamic enumeration of all mixed cells. Discrete & Computational
Geometry, 37(3):351–367, 2007.
[Oka97]
M. Oka. Non-degenerate complete intersection singularity. Hermann, Paris, 1997.
[Pao92a]
C. Pao. Nonlinear parabolic and elliptic equations. Plenum
Press, New York, 1992.
[Pao92b]
C.V. Pao. Nonlinear parabolic and elliptic equations. Plenum
Press, 1992.
[Par95]
L.M. Pardo. How lower and upper complexity bounds meet in
elimination theory. In G. Cohen, M. Giusti, and T. Mora, editors, Applied Algebra, Algebraic Algorithms and Error Correcting Codes, Proceedings of AAECC–11, volume 948 of Lecture
Notes in Comput. Sci., pages 33–69, Berlin, 1995. Springer.
[Par00]
L.M. Pardo. Universal elimination requires exponential running
time. In A. Montes, editor, Computer Algebra and Applications,
Proceedings of EACA–2000, Barcelona, Spain, September 2000,
pages 25–51, 2000.
[PR11]
Adrien Poteaux and Marc Rybowicz. Complexity bounds for
the rational newton-puiseux algorithm over finite fields. Applicable Algebra in Engineering, Communication and Computing,
22(3):187–217, 2011.
[PR12]
Adrien Poteaux and Marc Rybowicz. Good reduction of puiseux
series and applications. J. Symb. Comput., 47(1):32–63, 2012.
[PS93]
P. Pedersen and B. Sturmfels. Product formulas for resultants
and Chow forms. Math. Z., 214(3), 1993.
[PS03]
P. Philippon and M. Sombra. Hauteur normalisée des variétés
toriques projectives. eprint math.NT/0406476, 38pp., 2003.
[PS04]
L.M. Pardo and J. San Martín. Deformation techniques to
solve generalized Pham systems. Theoret. Comput. Sci., 315(2–
3):593–625, 2004.
182
BIBLIOGRAPHY
[PS05]
P. Philippon and M. Sombra. Géométrie diophantienne et variétés toriques. C. R. Math. Acad. Sci. Paris, 340:507–512, 2005.
[PS08]
P. Philippon and M. Sombra. Hauteur normalisée des variétés
toriques projectives. J. Inst. Math. Jussieu, 7(2):327–373, 2008.
[Roj99]
J.M. Rojas. Solving degenerate sparse polynomial systems
faster. J. Symbolic Comput., 28(1/2):155–186, 1999.
[Roj00]
J.M. Rojas. Algebraic geometry over four rings and the frontier of tractability. In J. Denef et al., editor, Proceedings of
a Conference on Hilbert’s Tenth Problem and Related Subjects
(University of Gent, November 1–5, 1999), volume 270 of Contemp. Math., pages 275–321, Providence, RI, 2000. Amer. Math.
Soc.
[Roj03]
J.M. Rojas. Why polyhedra matter in non–linear equation solving. In Proceedings conference on Algebraic Geometry and Geometric Modelling (Vilnius, Lithuania, July 29–August 2, 2002),
volume 334 of Contemp. Math., pages 293–320, Providence, RI,
2003. Amer. Math. Soc.
[Rou97]
F. Rouillier. Solving zero–dimensional systems through rational
univariate representation. Appl. Algebra Engrg. Comm. Comput., 9(5):433–461, 1997.
[Sai83]
R. Saigal. A homotopy for solving large, sparse and structured
fixed–point problems. Math. Oper. Res., 8:557–578, 1983.
[Sch03]
E. Schost. Computing parametric geometric resolutions. Appl.
Algebra Engrg. Comm. Comput., 13:349–393, 2003.
[SGKM95]
A.A. Samarskii, V.A. Galaktionov, S.P. Kurdyumov, and A.P.
Mikhailov. Blow–up in quasilinear parabolic equations, volume 19 of de Gruyter Expositions in Mathematics. Walter de
Gruyter, Berlin, 1995.
[Sha94]
I.R. Shafarevich. Basic Algebraic Geometry: Varieties in Projective Space. Springer, Berlin Heidelberg New York, 1994.
BIBLIOGRAPHY
183
[SS93a]
M. Shub and S. Smale. Complexity of Bézout’s theorem I: Geometric aspects. J. Amer. Math. Soc., 6(2):459–501, 1993.
[SS93b]
M. Shub and S. Smale. Complexity of Bézout’s theorem II:
Volumes and probabilities. In F. Eyssette and A. Galligo, editors, Computational Algebraic Geometry, volume 109 of Progr.
Math., pages 267–285, Boston, 1993. Birkhäuser Boston.
[SS93c]
M. Shub and S. Smale. Complexity of Bézout’s theorem III:
Condition number and packing. J. Complexity, 9:4–14, 1993.
[SS94]
M. Shub and S. Smale. Complexity of Bézout’s theorem V:
Polynomial time. Theoret. Comput. Sci., 133:141–164, 1994.
[SS96a]
J. Sabia and P. Solernó. Bounds for traces in complete intersections and degrees in the Nullstellensatz. Appl. Algebra Engrg.
Comm. Comput., 6(6):353–376, 1996.
[SS96b]
M. Shub and S. Smale. Complexity of Bézout’s theorem IV:
Probability of success. SIAM J. Numer. Anal., 33:141–164,
1996.
[Sto00]
A. Storjohann. Algorithms for matrix canonical forms. PhD
thesis, ETH, Zürich, Switzerland, 2000.
[VA85]
A. Varchenko V. Arnold, S. Gusein-Zade. Singularities of Differentiable Maps, volume I. Birkhäuser, Boston, 1985.
[Ver09]
Jan Verschelde. Polyhedral methods in numerical algebraic geometry. In D.J. Bates, G. Besana, and S. Di Rocco, editors,
Interactions of Classical and Numerical Algebraic Geometry,
volume 496 of nteractions of Classical and Numerical Algebraic
Contemporary Mathematics, pages 243–263. AMS, 2009.
[VGC96]
J. Verschelde, K. Gatermann, and R. Cools. Mixed volume
computation by dynamic lifting applied to polynomial system
solving. Discrete Comput. Geom., 16(1):69–112, 1996.
[Vog84]
W. Vogel. Results on Bézout’s theorem, volume 74 of Tata Inst.
Fundam. Res. Lect. Math. Tata Inst. Fund. Res., Bombay, 1984.
184
BIBLIOGRAPHY
[VVC94]
J. Verschelde, P. Verlinden, and R. Cools. Homotopies exploiting Newton polytopes for solving sparse polynomial systems.
SIAM J. Numer. Anal., 31(3):915–930, 1994.
[vzG86]
J. von zur Gathen. Parallel arithmetic computations: a survey. In J. Gruska, B. Rovan, and J. Wiedermann, editors, Proceedings of the 12th International Symposium on Mathematical
Foundations of Computer Science, Bratislava, Czechoslovakia,
August 25–29, 1996, volume 233 of Lecture Notes in Comput.
Sci., pages 93–112, Berlin, August 1986. Springer.
[vzGG99]
J. von zur Gathen and J. Gerhard. Modern computer algebra.
Cambridge Univ. Press, Cambridge, 1999.
[Wal50]
R.J. Walker. Algebraic Curves. Dover Publications Inc., New
York, 1950.
[Wal99]
P.G. Walsh. On the complexity of rational Puiseux expansions.
Pacific J. Math., 188(2):369–387, 1999.
[Wal00]
P.G. Walsh. A polynomial–time complexity bound for the computation of the singular part of a Puiseux expansion of an algebraic function. Math. Comp., 69(231):1167–1182, 2000.
[Zip93]
R. Zippel. Effective Polynomial Computation, volume 241 of
Kluwer Internat. Ser. Engrg. Comput. Sci. Kluwer Acad. Publ.,
Dordrecht, 1993.
Firma de la Tesis
Director de Tesis
Tesista
Firma
Firma
Fecha
Fecha

Biblioteca Digital | FCEN-UBA | Waissbein, Ariel. 2013 "Algoritmos

Transcription

Similar documents

A family of newforms

Pedro Fernández began his international career when he was

IMA Journal of Numerical Analysis (1992) 12, 387

PDF - Visual Optics Institute

Dedicado a todas las personas que creen en la musica

Specification Guide - Launch Distribution

Beerbaum Art International diversity

el libro de colorear, Jaime, el Dulce Guerrero

Numerical CFD comparison of Lillgrund employing RANS

TODAY`S READINGS First Reading — When the wicked turn away